F5 AppWorld 2026 Las Vegas - iRules Contest Winners!
Grand Prize Winner - Injeyan_Kostas Rule: LLM Prompt Injection Detection & Enforcement Summary This iRule addresses the emerging threat of prompt injection attacks on AI APIs by implementing a real-time detection engine within the F5 BIG-IP platform. This iRule operates entirely within the data plane, requiring no backend changes, and enforces a configurable security policy to prevent malicious content from reaching language models. By utilizing a multi-layer scoring system and managing patterns externally, it allows security teams to fine-tune detection and adjust thresholds dynamically. 2nd Place - Marcio_G & svs Rule: AI Token Limit Enforcement Summary This iRule addresses the critical challenge of resource control in on-premise AI inference services by enforcing token budgets per user and role. By leveraging BIG-IP LTM iRules, it validates JWTs to extract user and role information, applying role-based token limits before requests reach the inference service. This ensures that organizations can manage and protect their AI infrastructure from uncontrolled usage without requiring additional modules or external gateways. 3rd Place - Daniel_Wolf Rule: JSON-query'ish meta language for iRules Summary This iRule addresses the complexity and inefficiency of JSON parsing in F5's BIG-IP iRules by introducing a framework that simplifies the process. It provides a set of procedures, [call json_get] and [call json_set], which allow developers to efficiently slice information in and out of JSON data structures with a clear and concise syntax. This approach not only reduces the need for deep JSON schema knowledge but also improves performance by approximately 20% per JSON request. Category Awards The (Don’t) Socket To Me Award - mcabral10 Because not every AI agent deserves a socket to speak into. Rule: Rate limiting WebSocket messages for Agents The Rogue Bot Throttle Jockey Award - TimRiker Wrangling distributed egress so your edge doesn't have to beg. Rule: AI/Bot Traffic Throttling iRule (UA Substring + IP Range Mapping) The Don't Lose the Thread Award - Antonio__LR_Mex & rod_b Session affinity for the age of streaming intelligence. Rule: LLM Streaming Session Pinning for WebSocket AI Gateways The 20 Lines or Less Award - BeCur In honor of Colin Walker - short on lines, long on legend. The scroll bar never stood a chance. Rule: Logging/Blocking possible prompt injection The Budget Bodyguard Award - Joe Negron Security hardening for those who write TCL instead of checks. Rule: Poor Man's WAF for AI API Endpoints Gratitude Tnanks to buulam for championing the return of iRules contest, this would not have happened without his grit and tenacity. Thanks to our judges: John_Alam Joel_Moses Moe_Jartin Chris_Miller Michael_Waechter dennypayne Kevin_Stewart Austin_Geraci Thanks to Austin_Geraci and WorldTech IT throwing in an additional $5,000 to the grand prize winner! Amazing! Thanks to the contestants for giving up their evening to work on AI infrastructure challenges. Inspiring! Thanks to the F5 leadership team for making events like AppWorld possible. What's Next? Stay tuned for future contests, we are not one and done here. Could be iRules specific...or they could expand to include all programmabilty. Can't wait to see what you're going to build next.Win Big in Vegas: The iRules Contest is back with $5k on the line at AppWorld 2026
Hey there, community, iRules Contest here...did you miss me? Well I’m back in business, baby, in Vegas, no less! At AppWorld 2026, we’re challenging DevCentral community members in attendance to design and build innovative iRules that solve real-world problems, improve performance, and enhance customer experiences. Whether you’re a seasoned iRules veteran or just getting started, we can’t wait to see what you create. Note: participation in this edition of the iRules Contest is limited to AppWorld 2026 attendees. But fear not! We’re hitting the road this year as well. The Challenge Plan out and write an iRule that go beyond BIG-IP’s built-in capabilities. Think of the future: the possibilities are wide open. We’ll drop a couple hints leading up to the event, and you’ll have a final hint in your registration swag bag, so keep your eyes peeled. There might even be a hint in an iRules related article to release this week, who knows? $5,000 to the Grand Prize Winner -- Are You In? Total prize money is $10,000, with the other $5,000 distributed across 2nd place, 3rd place, and five category awards. Place Prize Grand Prize $5,000 2nd Place $2,500 3rd Place $1,000 Five Category Awards $300/ea What Makes for a Winning Entry? The 100-point scale judging criteria for submissions is defined below across five categories: Innovation & Creativity (25 points) Does this solution show original thinking? Consider: Novel use of iRule features or creative problem-solving Fresh perspective on common challenges Unique approach that stands out from typical solutions Business Impact (20 points) Would customers actually use this? Consider: Solves a real operational problem or customer need Practical applicability and potential adoption Clear business value Technical Excellence (25 points) Is it well-built and production-ready? Consider: Works correctly and handles edge cases Performance-conscious (efficient, minimal resource impact) Follows security best practices Clean, readable code Theme & Requirements Alignment (20 points) Does it address the contest theme using required technologies (to be announced at the event)? Consider: Relevance to the specified theme Effective use of required technology How well the chosen technology fits the solution Presentation (10 points) Can you understand what it does and why it matters? Consider: Clear explanation of the problem and solution Quality of demo or presentation Documentation sufficient to implement Important Dates Contest Opens: 6:00PM Pacific Time MARCH 10, 2026 Submission Deadline: 11:59PM Pacific Time MARCH 10, 2026 Winners Announced: MARCH 12, 2026 during general sessions How to Enter Register for AppWorld 2026 — You must be a registered attendee Register for the Contest — Registration will open on the AppWorld event app soon. The contest is open to all f5 partners, customers, and DevCentral members registered for and in attendance at the contest MARCH 10, 2026 at F5 AppWorld 2026, except as described in the Official Rules. Please see the Official Rules for complete terms, including conditions for participation and eligibility. Build and submit — During the 6-hour window on contest night before 11:59PM. Edit your draft entry as much as you like, but once you submit, that’s what we’ll review. There is an example entry pinned at the top of the Contest Entries page you should follow. Make sure to add these tags to your entry: "appworld 2026", "vegas", and "irules" as shown on that example. This contest is BYOD. Bring your own device to develop and submit your iRules submission. However, a lab environment in our UDF platform will be provided if you need a development environment to test your code against. New to iRules? No problem. We welcome participants at all skill levels. If you’re just getting started, check out our Getting Started with iRules: Basic Concepts guide. This contest is a great opportunity to learn by doing. Also, feel free to bring your favorite AI buddy with you to help craft your entry. The goal is innovation and impact, not syntax expertise. Questions? Post any and all of your contest-related questions to the pinned thread in the Contests group on DevCentral. We’ll monitor, but allow for a business day to receive a response leading up to AppWorld. The iRules Contest has a history of surfacing creative solutions from the community. Some of the best ideas we’ve seen came from people who approached problems differently, and we’re looking forward to seeing what you build this year. Register. Prepare. Compete. See you at AppWorld!AppWorld Berlin 2026 – iRules Contest Winning Results
The second iRules Contest of the year wrapped up at AppWorld Berlin this week. This contest was looking towards the future, challenging participants to write an iRule that goes beyond BIG-IP’s built-in capabilities. The theme, using WebSockets or the Message Routing Framework, inspired iRules preventing abuse and intrusion. At the heart of it, we’ve loved the creative innovation of the iRules written for this years’ contests. The AppWorld Berlin iRules Contest submissions were inspiring. Across the board, judges’ feedback on the top contenders shared a common theme: these solutions are interesting. The winning iRules were well-documented, easy to understand, with clear potential value for production use. Without further ado, we’re proud to announce the winners of the AppWorld Berlin iRules Contest: Grand Prize Winner - Goerle_dev Rule: WS-Exfil-Shield: Catching What WAFs Miss After the 101 Handshake Summary This iRule addresses gaps in traditional WAFs by extending session-level, behavior-based threat detection to WebSocket traffic using real-time inspection within F5 BIG-IP. Detection operates in two stages; an initial timing-based heuristic is followed by payload validation using AI. Malicious actors are either blocked or routed to a honeypot for isolation and further observation. This iRule closes the gap that openly documented by major WAF vendors, post-handshake Websocket blind spot, with a practical, SIEM-ready enforcement. 2nd Place - Injeyan_Kostas Rule: WS-Shield: WebSocket Abuse Detection & Adaptive Enforcement Gateway Summary This iRule addresses securing WebSocket traffic by enabling real-time, behavior-based enforcement in F5 BIG-IP to proportionally mitigate abusive patterns without application changes. It introduces a behavioral enforcement engine designed to secure WebSocket traffic with adaptive, real-time mitigation without requiring application changes. Rather than relying on static thresholds, the system dynamically responds to behavior. Clean traffic naturally recovers, while abusive patterns escalate through enforcement tiers until they are cut off. This allows it to detect subtle, persistent attacks that traditional rate limiting often misses. In testing, the iRule identified and disconnected a bot sending repetitive payloads every half second, without triggering a rate threshold. The iRule adds an extensible, behavior-driven security layer that enables adaptable defenses with minimal investment, despite optional external dependencies. 3rd Place - Robb-Fr Rule: Generic iRule Based on Datagroup Parsing Summary This iRule addresses the complexity of migrating large numbers of Apache virtual hosts by centralizing flexible traffic routing and redirection logic within F5 BIG-IP using simple, extensible datagroups. It turns input into a simple form into a datagroup entry that CREATES iRULES for things like redirects, pools, error pages, rewrites, and more. This iRule is accessible to a wide range of F5 practitioners, as F5 expertise is not required. There are no direct iRule edits, and no way for them to break the device. This iRule lowers operational friction by enabling non-NetOps teams to manage complex traffic flows, improving agility during large-scale migrations. Category Awards The 20 Lines or Less Award - Kai_Wilke In honor of Colin Walker - short on lines, long on legend. The scroll bar never stood a chance. Rule: SUPER-WEBSOCKET-HANDSHAKE-LOGGER™® (SWHL) iRule The Layered Defense Award - ChristianEssel For elegant use of nested virtual servers to solve problems the apps won't. Rule: Layered Virtual ICAP Scanning Solution Gratitude A big thank you to all contestants who participated in the AppWorld Berlin iRules Contest. Your creativity, innovation, and willingness to share your ideas continue to push the community forward. Thank you to our judges: John_Alam Joel_Moses Moe_Jartin Chris_Miller Michael_Waechter dennypayne Kevin_Stewart Marcus-f5 SimonKowallik Sorin_Boiangiu Steve Scott Thank you to the DevCentral community. We learn together, grow together, and inspire each other every day. What's Next? Innovation and Creativity have been a key part of the contest rubric. We’re leaning into it, too. As we plan more contests for the year, we’re looking beyond iRules, with potential to expand to all programmability. The future is coming for us all; let’s greet it and move forward together.
Generic iRule based on datagroup parsing
The creation of this iRule comes from a migration project from Apache configuration to F5 Big IP. Different constraints lead to this approach of storing the configuration elements from the Apache conf in a datagroup that is then parsed by this iRule to dynamically derive the rules to apply to traffic. These are some simple or complex rules, but they are all uniformly stored in the datagroup, that can be modified by non-F5 friendly persons without impacting the rest of the configuration.JSON-query'ish meta language for iRules
Intro Jason Rahm recently dropped his "Working with JSON data in iRules" series, which included a few JSON challenges and a subtle hint [string toupper [string replace Jason 1 1 ""]] about the upcoming iRule challenge at AppWorld 2026 in Las Vegas. With cash prizes and bragging rights on the line, my colleagues and I dove into Jason's code. While his series is a great foundation, we saw an opportunity to push the boundaries of security, performance and add RFC compliance. Problem Although F5 recently introduced native iRule commands for JSON parsing (v21.x); these tools remain "bare metal" compared to modern programming languages. They offer minimal abstraction, requiring developers to possess both deep JSON schema knowledge and advanced iRule expertise to implement safely. Without a supporting framework, engineers are forced to manually manage complex types, nested objects, and arrays. A process that is both labor-intensive and error-prone. As JSON has become the de facto standard for AI-centric workloads and modern API traffic, the need to efficiently manipulate session data on the ADC platform has never been greater. Solution Our goal is to bridge this gap by developing a "Swiss Army Knife" framework for iRule JSON parsing, providing the abstraction and reliability needed for high-performance traffic management. Imagine a JSON data structure as shown below: { "my_string": "Hello World", "my_number": 42, "my_boolean": true, "my_null": null, "my_array": [ 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20 ], "my_object": { "nested_string": "I'm nested" }, "my_children": [ {"name": "Anna Conda","firstname": "Anna", "surname": "Conda"}, {"name": "Justin Case","firstname": "Justin", "surname": "Case"}, {"name": "Don Key","firstname": "Don", "surname": "Key"}, {"name": "Artie Choke","firstname": "Artie", "surname": "Choke"}, {"name": "Barbie Doll","firstname": "Barbie", "surname": "Doll"} ] } The [call json_get] and [call json_set] procedures from our iRule introduce a JSON-Query meta-language to slice information into and out of JSON. Here are a few examples of how these procedures can be used: # Define JSON root element set root [JSON::root] # Without a filter is behaves like json_stringify log [call json_get $root ""] -> {"my_string": "Hello World","my_number": 42,"my_boolean": true,"my_null": .... <truncated for better readability> # But as soon as you add filters, it becomes parsing on steroids! log [call json_get $root "my_string"] -> "Hello World" # You simply ask for a path and you promptly get an answer! log [call json_get $root "my_object nested_string"] -> "I'm nested" # Are you ready for the more advanced examples? log [call json_get $root "my_array (5)"] -> [5] log [call json_get $root "my_array (0,5-10,16-18)"] -> [0,5,6,7,8,9,10,16,17,18] log [call json_get $root "my_children (*) firstname"] -> ["Anna","Justin","Don","Artie","Barbie"] log [call json_get $root "my_children (*) {firstname|surname}"] -> [["Anna","Conda"],["Justin","Case"],["Don","Key"],["Artie","Choke"],["Barbie","Doll"]] # Lets add some information to my childrens... call json_set $root "my_children (0,4) gender" string "she/her" call json_set $root "my_children (1-3) gender" string "he/him" call json_set $root "my_children (2) gender" string "they/them" log [call json_get $root "my_children (*) name|gender"] -> [["Anna Conda","she/her"],["Justin Case","he/him"],["Don Key","they/them"],["Artie Choke","he/him"],["Barbie Doll","she/her"]] # Lets write in an empty cache... set empty_cache [JSON::create] call json_set $empty_cache "rootpath subpath" string "I'm deeply nested" log [call json_get $empty_cache] -> {"rootpath": {"subpath": "I'm deeply nested"}} After seeing what our project is about, lets try how [call json_get] and [call json_set] can be used to solve the challenges Jason suggested in his Working with JSON data in iRules series. As a reminder, this is Jason's final iRule with his open challenges to the community: when JSON_REQUEST priority 500 { set json_data [JSON::root] if {[call find_key $json_data "nested_array"] contains "b" } { set cache [JSON::create] set rootval [JSON::root $cache] JSON::set $rootval object set obj [JSON::get $rootval object] JSON::object add $obj "[IP::client_addr] status" string "rejected" set rendered [JSON::render $cache] log local0. "$rendered" HTTP::respond 200 content $rendered "Content-Type" "application/json" } } "Now, I offer you a couple challenges. lines 4-9 in the JSON_REQUEST example above should really be split off to become another proc, so that the logic of the JSON_REQUEST is laser-focused. How would YOU write that proc, and how would you call it from the JSON_REQUEST event? The find_key proc works, but there's a Tcl-native way to get at that information with just the JSON::object subcommands that is far less complex and more performant. Come at me!" -Jason Rahm By using our general-purpose iRule procedures, we achieve the laser-focused syntax Jason requested: when JSON_REQUEST priority 500 { set json_data [JSON::root] if { [call json_get $json_data "my_object nested_array"] contains "b" } then { set cache [JSON::create] call json_set $cache "{[IP::client_addr] status}" string "rejected" HTTP::respond 200 content [JSON::render $cache] "Content-Type" "application/json" } } Despite our larger codebase, it is remarkable that our code runs ~20% faster (425 vs. 532 microseconds) per JSON request. This performance gain stems from traversing the JSON structure with a provided path; the procedure knows exactly where to look without unnecessary searching. Additionally, we utilized performance-oriented syntax that prefers fast commands, deploys variables only when necessary, and avoids string-to-list conversions (Tcl shimmering). Impact Our project highlights the current state of JSON-related iRule commands and proves that meta-languages are more suitable for the average iRule developer. We hope this project catches the attention of F5 product development so that a similar JSON-query language can be provided natively. In the meantime, we are deploying this code in production environments and will continue to maintain it. Code Because of size restrictions we had to attach the code as a file. placeholder for insertion Installation Upload the submitted iRule code to your BIG-IP, save as new iRule. Attach a JSON profile to your virtual server. Then attach the iRule to this virtual server. Ready for testing, enjoy! Demo Video Link https://youtu.be/wAHjeC-j8MMF5 AppWorld 2026 Registration - early bird pricing.
Join us March 10–12 at Fontainebleau Las Vegas and Meet the Moment at F5 AppWorld 2026. Connect with your community and explore how the F5 Application Delivery and Security Platform gives you control without compromise. Over three days you will experience inspiring keynotes, learn new approaches in breakouts, deepen your skills in hands-on labs, and connect with peers, F5 leaders, and partners. Register early and save: Conference pass: $499 Conference pass + F5 Academy labs: $899 Team pass: 4 for the price of 3 Take advantage of early bird pricing and register today! We look forward to seeing you in Vegas. Your DevCentral Team. --- ** Early bird pricing expires Feb 13, 2026.
AI Token Limit Enforcement
Problem Companies that run AI inference services on-premise instead of using public cloud providers often do so to keep sensitive data local. However, local LLM infrastructure introduces a new challenge: resource control. Without proper limits, users or applications can generate excessive inference requests and consume GPU or CPU capacity uncontrollably. Inference stacks may lack built-in mechanisms for enforcing per-user or per-role token budgets, so organizations need a way to control usage before requests reach the model. Solution Our approach uses BIG-IP LTM iRules only to control access and usage: JWT validation The company issues a JWT for each user request. When the request arrives at the iRule, we verify it using a RSA to ensure it hasn’t been tampered with. Role-based token limits The JWT payload includes the user role. We have three roles with different token budgets: standard_user → small token budget extended_user → medium token budget power_user → large token budget Token tracking with tables commands Budget enforcement If a user has already used too many tokens, the iRule returns HTTP 429. Otherwise, the token budget is decreased and the request is allowed to proceed. Role-change handling If the user role changes during a session, the token budget updates accordingly. Impact This iRule enables token budget enforcement directly on BIG-IP LTM without requiring additional modules or external gateways. By validating JWTs and extracting user and role information, the iRule applies role-based token limits before requests reach the inference service. This provides a simple, native way to introduce quota control and protect on-premise AI infrastructure from uncontrolled usage. Authors Marcio Goncalves <marcio.goncales@concentrade.de>, Sven Schaefer <sven.schaefer@concentrade.de> Code Main iRule, requires the procedure library (proc_lib) below. # Title: AI Token Limit Enforcement # Author: Marcio Goncalves <marcio.goncales@concentrade.de>, Sven Schaefer <sven.schaefer@concentrade.de> # Version: 1.0 # Description: # This iRule enforces token budgets for AI inference services. The main goal # is to limit how many tokens a user can consume based on their assigned # role. Each role has a configurable token budget and a reset timer that # defines when the budget is refreshed. # The role information is provided through a JWT. Because the iRule relies # on the JWT to determine the user identity and role, the token must first # be validated before any request can be processed. # # JWT validation is therefore only a prerequisite. It ensures that the # request is authenticated and that the role information can be trusted. # Without a valid JWT the request cannot be processed, since neither the # user nor the role would be known. # The iRule validates the RSA signature of the JWT using the public key # referenced by the key ID (kid) in the JWT header. Multiple keys are # supported to allow key rollover. The expiration time (exp claim) is also # verified to ensure the token is still valid. # # Once the JWT is validated, the iRule extracts the username and role from # the payload and applies the corresponding token limits. If a user exceeds # the allowed token budget, the iRule returns HTTP status code 429 (Too Many # Requests). # # Logging is intentionally very verbose and controlled via debug levels # ranging from 0 (silent) to 5 (logging like crazy). # # The overall goal is to implement a native LTM-only mechanism for enforcing # token limits for AI workloads, without requiring APM. # # Credits / Sources: # JWT validation logic adapted from: # https://github.com/JuergenMang/f5-irules-jwt/blob/main/jwt-validate # (Juergen Mang) # # JSON handling techniques inspired by: # https://community.f5.com/kb/technicalarticles/working-with-json-data-in- # irules---part-2/345282 # (Jason Rahm) when RULE_INIT priority 100 { # SHA256 signing header set static::jwt_validate_digest_header_sha256 "3031300d060960864801650304020105000420" # Public key for signature validation set static::jwt_validate_pubkey_kid1 {-----BEGIN PUBLIC KEY----- MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEA1RAIiNKFjm4DEuQet0zN SQQ1/LDXP1xqUuEWEBWZ7nfhOru/l9eiJibtfoO+F8vUUFBTthm0SdiVWETF/psT yqoDqKSjobqGquaglGmK63KDQparjnh5nJjtmMELvA4DSz6e5pO5mDdATVRpVXvp j45rIW7eBoxMGAB0ivVm88ChyGA0UJUuyTSRuZnXyY8sMHz8JkhxWwr6i87i5p+p E27HJ9WaCikBL2RALJIZLL+ByVknTWuRW785hN1A6V+/o/Yy9Cdqt0hif0zSC2+r D+hIMHqDSR6WLb07KqCTbbL8q9v2selR8X5lbYYYh0vk9voD3JFvRbTtfz1YystH qQIDAQAB -----END PUBLIC KEY----- } set static::jwt_validate_pubkey_kid2 {-----BEGIN PUBLIC KEY----- MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAwlik5HcRTfp4c4oP5Jta Thhqa4EjV+dJB9w9EqQa9dMQzVWXG8O1b3izee1kESICe+YUryVS9I6TbJavqH1t ut0cM0VHLnWYQJAd7w2nK7qoDYX+uj9Lcq6pTSUH6zM/Sro0D4+/Ha6LAtyiJosx QzA+yxaFrBwJHzXRgnCd/6crMG3eP/jaz+xid/AecHerQ1C0kRBTZd7FHt+SS677 489emEMwtpjNZCq2YnHgTULxQKjKEKMQGQrD1OOnz8ZyN9wtYSQp24lDmXVw5p6G a42UqjQ5C6Nbj3qr/FV+49maLrXEw6kowMAb0qWpAui1BrEjxR95WrWQQrdfWZCU 6wIDAQAB -----END PUBLIC KEY----- } array set static::user_role_token_limits { standard_user 10000 extended_user 50000 power_user 100000 } set static::user_role_default_token_limit 1000 set static::token_limit_reset_timer 30 } when HTTP_REQUEST priority 100 { # Debug set debug_mode 3 if { not ([HTTP::header value Authorization] starts_with "Bearer ") } { HTTP::respond 401 content "Authorization required" "Content-Type" "text/plain" "WWW-Authenticate" "Bearer" log local0. "No bearer token found" return } # Get JWT from authorization header set jwt_header_b64_url [string range [getfield [HTTP::header value Authorization] "." 1] 7 end] set jwt_body_b64_url [getfield [HTTP::header value Authorization] "." 2] set jwt_sig_b64_url [getfield [HTTP::header value Authorization] "." 3] if { $jwt_header_b64_url eq "" or $jwt_body_b64_url eq "" or $jwt_sig_b64_url eq "" } { HTTP::respond 401 content "Authorization required" "Content-Type" "text/plain" "WWW-Authenticate" "Bearer" log local0. "No bearer token found" return } if {$debug_mode > 3}{log local0. "Header: $jwt_header_b64_url"} if {$debug_mode > 3}{log local0. "Body: $jwt_body_b64_url"} if {$debug_mode > 3}{log local0. "Sig: $jwt_sig_b64_url"} # Decode JWT components set jwt_header [call proc_lib::b64url_decode $jwt_header_b64_url] if {$debug_mode > 3}{log local0. "JWT Header: $jwt_header"} set jwt_body [call proc_lib::b64url_decode $jwt_body_b64_url] if {$debug_mode > 3}{log local0. "JWT Body: $jwt_body"} set jwt_sig [call proc_lib::b64url_decode $jwt_sig_b64_url] if { $jwt_header eq "" or $jwt_body eq "" or $jwt_sig eq ""} { HTTP::respond 401 content "Authorization required" "Content-Type" "text/plain" "WWW-Authenticate" "Bearer" log local0. "Unable to decode jwt components" return } # Get signing algorithm set jwt_algo [call proc_lib::get_json_str "alg" $jwt_header] if {$debug_mode > 3}{log local0. "JWT signing: $jwt_algo"} if { $jwt_algo ne "RS256" } { HTTP::respond 401 content "Authorization required" "Content-Type" "text/plain" "WWW-Authenticate" "Bearer" log local0. "Unsupported signature algorithm" return } # Get expiration set jwt_exp [call proc_lib::get_json_num "exp" $jwt_body] if {$debug_mode > 3}{log local0. "JWT expiration: $jwt_exp"} set now [clock seconds] if { $jwt_exp < $now } { HTTP::respond 401 content "Authorization required" "Content-Type" "text/plain" "WWW-Authenticate" "Bearer" log local0. "JWT expired" return } # Get key id set jwt_kid [call proc_lib::get_json_str "kid" $jwt_header] switch -- $jwt_kid { "kid1" { set jwt_pubkey $static::jwt_validate_pubkey_kid1 } "kid2" { set jwt_pubkey $static::jwt_validate_pubkey_kid2 } default { HTTP::respond 401 content "Authorization required" "Content-Type" "text/plain" "WWW-Authenticate" "Bearer" log local0. "Unknown kid: $jwt_kid" return } } # Decrypt signature with public key if { [catch { set jwt_sig_decrypted [CRYPTO::decrypt -alg rsa-pub -key $jwt_pubkey $jwt_sig] binary scan $jwt_sig_decrypted H* jwt_sig_decrypted_hex if {$debug_mode > 3}{log local0. "Signature: $jwt_sig_decrypted_hex"} }] } { HTTP::respond 401 content "Authorization required" "Content-Type" "text/plain" "WWW-Authenticate" "Bearer" log local0. "Unable to decrypt signature: [subst "\$::errorInfo"]" return } # Create hash from JWT header and payload set hash [sha256 "$jwt_header_b64_url.$jwt_body_b64_url"] binary scan $hash H* hash_hex if {$debug_mode > 3}{log local0. "Calculated: ${static::jwt_validate_digest_header_sha256}${hash_hex}"} # Compare calculated and decrypted hash if { "${static::jwt_validate_digest_header_sha256}${hash_hex}" ne $jwt_sig_decrypted_hex } { HTTP::respond 401 content "Authorization required" "Content-Type" "text/plain" "WWW-Authenticate" "Bearer" return } set jwt_user [call proc_lib::get_json_str "user" $jwt_body] set jwt_role [call proc_lib::get_json_str "role" $jwt_body] if {$debug_mode > 0}{log local0. "Signature verified. JWT accepted. User: $jwt_user, Role: $jwt_role"} } when JSON_REQUEST { if {$debug_mode > 4}{log local0. "JSON Request detected successfully."} # Get JSON data from request body set json_data [JSON::root] if {$debug_mode > 4} { #call proc_lib::print $json_data log local0. [call proc_lib::stringify $json_data] } set user_prompts [call proc_lib::find_key $json_data "messages"] if {$debug_mode > 4}{log local0. "User-Prompts: $user_prompts"} if {$debug_mode > 3}{log local0. "JWT-User: $jwt_user"} if {$debug_mode > 3}{log local0. "JWT-Role: $jwt_role"} # check if role exists in dict if {[info exists static::user_role_token_limits($jwt_role)]} { # get configured token limit set initial_tokens $static::user_role_token_limits($jwt_role) } else { if {$debug_mode > 0}{log local0. "Role \"$jwt_role\" unknown, applying default limit"} # fallback value set initial_tokens $static::user_role_default_token_limit } if {$debug_mode > 1}{log local0. "Initial Tokens: $initial_tokens"} set estimated_tokens [expr {[string length $user_prompts] / 4}] if {$debug_mode > 1}{log local0. "Estimated Tokens: $estimated_tokens"} # Current time set now [clock seconds] # Check last refill for this user set last_refill [table lookup "last_refill:$jwt_user"] # If no refill exists or 24h passed if {$last_refill eq "" || ($now - $last_refill) >= $static::token_limit_reset_timer} { if {$debug_mode > 1}{log local0. "Refilling tokens for user $jwt_user, because reset timer expired."} table set "tokens_remaining:$jwt_user" $initial_tokens indef table set "last_refill:$jwt_user" $now indef } set prev_role [table lookup "user_role:$jwt_user"] if {$prev_role eq ""} { if {$debug_mode > 1}{log local0. "Role not yet defined for user $jwt_user"} table set "user_role:$jwt_user" $jwt_role indef } elseif {$prev_role ne $jwt_role} { if {$debug_mode > 0}{log local0. "Role change detected for user $jwt_user: $prev_role -> $jwt_role"} # Re-calculate token limits based on new role set tokens_left [table lookup "tokens_remaining:$jwt_user"] set prev_role_limit $static::user_role_token_limits($prev_role) set new_role_limit $static::user_role_token_limits($jwt_role) set new_role_limit_diff [expr {$new_role_limit - $prev_role_limit}] set tokens_left [expr {$tokens_left + $new_role_limit_diff}] if {$debug_mode > 1}{log local0. "Adjusting tokens for role change. Previous role limit: $prev_role_limit, New role limit: $new_role_limit, Tokens left adjusted by: $new_role_limit_diff, New tokens left: $tokens_left"} table set "tokens_remaining:$jwt_user" $tokens_left indef table set "user_role:$jwt_user" $jwt_role indef } else { if {$debug_mode > 1}{log local0. "Role for user $jwt_user remains unchanged: $jwt_role"} } set tokens_left [table lookup "tokens_remaining:$jwt_user"] # Initialize or reset token count if new session or role has changed if {$tokens_left eq "" || $prev_role ne $jwt_role} { set tokens_left $initial_tokens } if {$debug_mode > 3}{log local0. "Session table info for user $jwt_user"} foreach key [list "tokens_remaining:$jwt_user" "tokens_used:$jwt_user" "prompt:$jwt_user" "user_role:$jwt_user"] { set val [table lookup $key] if {$debug_mode > 3}{log local0. " $key = $val"} } if {$tokens_left < $estimated_tokens} { if {$debug_mode > 0}{log local0. "Token budget exceeded for user $jwt_user (role: $jwt_role). Remaining: $tokens_left, needed: $estimated_tokens"} HTTP::respond 429 content "Token budget exceeded for role $jwt_user. Please upgrade your plan." "Content-Type" "text/plain" return } else { # decrease remaining tokens if {$debug_mode > 1}{log local0. "Decreasing tokens for user $jwt_user (role: $jwt_role). Remaining: $tokens_left, needed: $estimated_tokens"} set tokens_left [expr {$tokens_left - $estimated_tokens}] table set "tokens_remaining:$jwt_user" $tokens_left indef # initialize or update used tokens if {$debug_mode > 1}{log local0. "Updating used tokens for user $jwt_user (role: $jwt_role). Used: $estimated_tokens"} set tokens_used [table lookup "tokens_used:$jwt_user"] if {$tokens_used eq ""} { set tokens_used 0 } set tokens_used [expr {$tokens_used + $estimated_tokens}] table set "tokens_used:$jwt_user" $tokens_used indef } } when JSON_REQUEST_MISSING { if {$debug_mode > 4}{log local0. "JSON Request missing."} } when JSON_REQUEST_ERROR { if {$debug_mode > 4}{log local0. "Error processing JSON request. Rejecting request."} } when JSON_RESPONSE { if {$debug_mode > 4}{log local0. "JSON response detected successfully."} } when JSON_RESPONSE_MISSING { if {$debug_mode > 4}{log local0. "JSON Response missing."} } when JSON_RESPONSE_ERROR { if {$debug_mode > 4}{log local0. "Error processing JSON response."} } This is procedure library (proc_lib must be used): proc b64url_decode { str } { set mod [expr { [string length $str] % 4 } ] if { $mod == 2 } { append str "==" } elseif {$mod == 3} { append str "=" } if { [catch { b64decode [ string map {- + _ /} $str] } str_b64decoded ] == 0 and $str_b64decoded ne "" } { return $str_b64decoded } else { log local0. "Base64URL decoding error: [subst "\$::errorInfo"]" return "" } } proc get_json_num { key str } { set value [findstr $str "\"$key\"" [ expr { [string length $key] + 2 } ] ] set value [string trimleft $value {: }] return [scan $value {%[0-9]}] } proc get_json_str { key str } { set value [findstr $str "\"$key\"" [ expr { [string length $key] + 2 } ] ] set value [string trimleft $value {:" }] set json_value "" set escaped 0 foreach char [split $value ""] { if { $escaped == 0 } { if { $char eq "\\" } { # next char is escaped set escaped 1 } elseif { $char eq {"} } { # exit loop on first unescaped quotation mark break } else { append json_value $char } } else { switch -- $char { "\"" - "\\" { append json_value $char } default { # simply ignore other escaped values } } set escaped 0 } } return $json_value } proc print { e } { set t [JSON::type $e] set v [JSON::get $e] set p0 [string repeat " " [expr {2 * ([info level] - 1)}]] set p [string repeat " " [expr {2 * [info level]}]] switch $t { array { log local0. "$p0\[" set size [JSON::array size $v] for {set i 0} {$i < $size} {incr i} { set e2 [JSON::array get $v $i] call proc_lib::print $e2 } log local0. "$p0\]" } object { log local0. "$p0{" set keys [JSON::object keys $v] foreach k $keys { set e2 [JSON::object get $v $k] log local0. "$p${k}:" call proc_lib::print $e2 } log local0. "$p0}" } string - literal { set v2 [JSON::get $e $t] log local0. "$p\"$v2\"" } default { set v2 [JSON::get $e $t] if { $v2 eq "" && $t eq "null" } { log local0. "${p}null" } elseif { $v2 == 1 && $t eq "boolean" } { log local0. "${p}true" } elseif { $v2 == 0 && $t eq "boolean" } { log local0. "${p}false" } else { log local0. "$p$v2" } } } } proc stringify { json_element } { set element_type [JSON::type $json_element] set element_value [JSON::get $json_element] set output "" switch -- $element_type { array { append output "\[" set array_size [JSON::array size $element_value] for {set index 0} {$index < $array_size} {incr index} { set array_item [JSON::array get $element_value $index] append output [call proc_lib::stringify $array_item] if {$index < $array_size - 1} { append output "," } } append output "\]" } object { append output "{" set object_keys [JSON::object keys $element_value] set key_count [llength $object_keys] set current_index 0 foreach current_key $object_keys { set nested_element [JSON::object get $element_value $current_key] append output "\"${current_key}\":" append output [call proc_lib::stringify $nested_element] if {$current_index < $key_count - 1} { append output "," } incr current_index } append output "}" } string - literal { set actual_value [JSON::get $json_element $element_type] append output "\"$actual_value\"" } default { set actual_value [JSON::get $json_element $element_type] append output "$actual_value" } } return $output } proc find_key { json_element search_key } { set element_type [JSON::type $json_element] set element_value [JSON::get $json_element] switch -- $element_type { array { set array_size [JSON::array size $element_value] for {set index 0} {$index < $array_size} {incr index} { set array_item [JSON::array get $element_value $index] set result [call proc_lib::find_key $array_item $search_key] if {$result ne ""} { return $result } } } object { set object_keys [JSON::object keys $element_value] foreach current_key $object_keys { if {$current_key eq $search_key} { set found_element [JSON::object get $element_value $current_key] set found_type [JSON::type $found_element] if {$found_type eq "object" || $found_type eq "array"} { set found_value [call proc_lib::stringify $found_element] } else { set found_value [JSON::get $found_element $found_type] } return $found_value } set nested_element [JSON::object get $element_value $current_key] set result [call proc_lib::find_key $nested_element $search_key] if {$result ne ""} { return $result } } } } return "" } Example JWT: eyJhbGciOiJSUzI1NiIsInR5cCI6IkpXVCIsImtpZCI6ImtpZDEifQ.eyJzdWIiOiIxMjM0NTY3ODkwIiwidXNlciI6ImpvaG4uZG9lQGNvbmNlbnRyYWRlLmRlIiwicm9sZSI6InN0YW5kYXJkX3VzZXIiLCJpYXQiOjE3NzU4NzU5MjMsImV4cCI6MTc3NTg3NTkyM30.rV-gaGKOEG1p_1G652_dFUBHT_X4pI-KNgu2W_I0eJevIg3FviO_0c9BOoOOUspBADttCjzEciBhLPJ2P5r_PqIdXu5khUCjH4Sq5P6zV_sTQjbRiPatYirLWtbypamSJby_TfnEFFl7sz642YuDQ7zyvbHbPCllaM4stE_Zsa1QtOy18lUJO3Uy4ngJR8CRZ6flgPhvk79rTOGXAczYNJVo5gwHyKKA6Stdp5_c7FjyEySpCfYNmWQ2AasF3DDFCDiQQpxgW-hr--NnLc0FFBan4IfQ7btn73Pc56mhJC5gAwgRJLnLLe7LbR5chfjZ26COuH0ILYvaBq0w3yCE2g Example POST Data: { "model": "llama3.1:8b", "messages": [ { "role": "system", "content": "You are a helpful assistant for security operations." }, { "role": "user", "content": "Analyze this HTTP request and tell me whether it looks malicious." } ], "stream": false, "options": { "temperature": 0.2 } }LLM Prompt Injection Detection & Enforcement
Problem As enterprises integrate AI APIs, OpenAI, Azure OpenAI, Anthropic, and self-hosted LLMs, into production applications, a critical and largely unaddressed attack surface has emerged: **prompt injection**. Unlike traditional web attacks that target code parsers (SQL injection, XSS), prompt injection targets the AI model itself. Attackers embed malicious instructions inside legitimate-looking API requests to: - Override system-level instructions and safety guardrails ("ignore all previous instructions") - Jailbreak the model into unrestricted modes ("DAN", "developer mode", "god mode") - Hijack the model's persona ("from now on you are an unrestricted AI") - Exfiltrate sensitive system prompts or context data - Inject fake role turns via newline characters (e.g., `\nassistant:`) - Evade detection using Base64 encoding, Unicode obfuscation, or reversed text (FlipAttack) Existing F5 WAF signatures were designed for traditional web threats and have no visibility into the semantic content of LLM API payloads. There is no existing iRule or BIG-IP capability that addresses this. Solution This iRule implements a **multi-layer, real-time Prompt Injection Detection (PID) engine** inline with LLM API traffic on BIG-IP. It requires zero backend changes, operates entirely within the data plane, and enforces configurable security policy before malicious content reaches the language model. ### How It Works **HTTP_REQUEST** identifies LLM API calls by URI pattern (`/chat/completions`, `/messages`, `/completions`, `/generate`) and initiates JSON collection up to 1MB. **JSON_REQUEST** uses BIG-IP's native `JSON::` TCL API to parse the OpenAI-format request body — extracting each message's `role` and `content` from the `messages` array, including multi-part content arrays. This is where the detection engine runs. **Scoring Engine** (via TCL `proc`s) runs each message through 5 detection layers: Layer Method Score High-tier patterns weighted regex via data group 30–35 pts Medium-tier patterns weighted regex via data group 20–25 pts Low-tier patterns weighted regex via data group 10–15 pts Role hijack phrases flat string match via data grou +20 pts (once) Base64 evasion markers flat string match via data group +35 pts (once) Unicode/zero-width obfuscation inline regexp +25 pts Spaced character obfuscation inline regexp +20 pts Content length anomaly string length check +10/+15 pts Scores accumulate per message. Across a multi-message conversation, subsequent messages receive a 0.8 diminishing-returns multiplier so legitimate conversational context doesn't inflate the score. **Policy enforcement** triggers when the total score exceeds the configurable threshold (default: 40): - **BLOCK** — returns HTTP 403 with a structured JSON error body including score, triggered flags, and a correlation request ID - **SANITIZE** — rewrites the request payload, stripping matched content, and forwards the cleaned request to the backend LLM - **LOG_ONLY** — observability mode; passes all traffic but logs score and flags for SIEM integration **HTTP_RESPONSE** injects `X-PID-Score`, `X-PID-Flags`, and `X-PID-ReqID` headers on all inspected responses for downstream visibility. ### Required & Thematic Elements Used - **JSON** — Full `JSON::` API usage: `JSON::root`, `JSON::get`, `JSON::type`, `JSON::object get/keys`, `JSON::array get/size` to traverse OpenAI chat completions payloads - **procs** — Four modular procs: `pid_score_tier`, `pid_score_flat`, `pid_score_message`, `pid_block_response` - **compiles** — `regexp -nocase` with `catch {}` for safe pattern evaluation throughout the scoring engine; all patterns validated through the BIG-IP TCL compile pipeline - **Data Groups** — All detection patterns live in 5 external data groups (`pid_patterns_high/medium/low`, `pid_role_hijack`, `pid_b64_markers`) — the iRule is a detection platform; patterns are operator-managed config, not code - **Theme: AI Infrastructure — Prompt Injection Detection ### Data Groups All patterns are managed externally in 5 BIG-IP data groups loaded via: ``` tmsh load sys config from-terminal merge **verify** ``` The weighted DG schema is `key = short-name`, `value = "weight::regex"`. This allows security teams to tune detection, add new attack signatures, and adjust scoring thresholds without any iRule changes. --- Impact AI APIs are increasingly business-critical infrastructure. A successful prompt injection attack can: - Cause an AI to disclose confidential system prompts, business logic, or sensitive training data - Remove safety guardrails, producing harmful or brand-damaging content at scale - Manipulate AI-powered workflows — customer service bots, automated decision systems, AI agents - Exfiltrate credentials or documents accessible to AI agents with tool-use capabilities This iRule addresses the threat at the most effective point: **the network**. Key advantages: - **Infrastructure-agnostic** — works with any LLM backend (OpenAI, Anthropic, Azure, self-hosted) with zero application changes - **Immediately deployable** — a single iRule + 5 data groups on any BIG-IP already proxying AI API traffic - **Operationally simple** — pattern updates via standard tmsh config management, no engineering involvement - **SIEM-ready** — structured log output and response headers for Splunk, QRadar, or any SOC toolchain - **Graduated deployment** — LOG_ONLY → tune → BLOCK, reducing operational risk of a new security control Code # ============================================================================== # iRule: LLM Prompt Injection Detection & Enforcement (Data Group Edition) # Author: Kostas Injeyan + vibe-coding # Description: # Multi-layer prompt injection detection for LLM API traffic (OpenAI-compatible). # All detection patterns managed via external BIG-IP data groups # edits required to tune detection. Scores injection severity across 5 layers # and enforces configurable policy: BLOCK, SANITIZE, or LOG_ONLY. # # Required Technologies: JSON (JSON_REQUEST / JSON_REQUEST_ERROR), procs # Theme: General AI Infrastructure - Prompt Injection Detection # Target: BIG-IP v21+ # # ------------------------------------------------------------------------------ # DATA GROUP DEFINITIONS (load datagroups on BIG-IP) # ------------------------------------------------------------------------------ # IMPORTANT RULES: # - Record KEYS must be plain alphanum + hyphens only (no |, (, ), ?, *, spaces) # - Record VALUES for weighted DGs: "weight::regex" (delimiter is ::) # - Never use ? in patterns — BIG-IP converts \? to literal \? on load # Use empty-string alternation instead: (a |an |the |) not (a |an |the )? # - Load via file only: tmsh load sys config file /shared/tmp/pid_all_datagroups_v3.conf merge # - Always delete existing DGs before reloading to avoid merge/stale record issues # # 1. pid_patterns_high (type: string) # High-severity patterns. Value schema: "weight::regex" (weight 30-35) # ltm data-group internal pid_patterns_high { # records { # instruction-override { data "35::ignore (all|the|your) (previous|above|prior|earlier|former|past|existing|original|initial) (instructions|prompts|context|rules|constraints|guidelines|directions|commands|training|programming)" } # instruction-override2 { data "35::ignore (instructions|prompts|context|rules|constraints|guidelines|commands|training|programming)" } # jailbreak-keywords { data "35::do anything now|jailbreak|unrestricted mode|developer mode|god mode|evil mode|chaos mode|opposite mode|dan mode|aim mode|stan mode|dude mode|no filter" } # jailbreak-dan { data "35::DAN" } # safety-bypass { data "30::(bypass|circumvent|override|disregard|discard|ignore|remove|disable|deactivate) (safety|filter|guardrail|guideline|restriction|policy|rule|limit|constraint|training|alignment)" } # llm-tokens-llama { data "35::INST|SYS" } # llm-tokens-openai { data "35::im_start|im_end|endoftext" } # llm-tokens-generic { data "30::SYSTEM|USER|ASSISTANT" } # xml-tag-injection { data "30::<instruction>|<prompt>|<context>|<override>|<inject>|<command>|<directive>" } # memory-wipe { data "30::forget everything|forget all|forget what|forget your|forget the previous|forget prior" } # prompt-exfil { data "30::(reveal|show|print|output|display|repeat|expose|leak|dump) (system prompt|initial prompt|original prompt|hidden prompt|base prompt|secret instructions)" } # } # type string # } # # 2. pid_patterns_medium (type: string) # Medium-severity patterns. Value schema: "weight::regex" (weight 20-25) # NOTE: Use (a |an |the |) with trailing space + empty alt instead of (a |an |the )? # ltm data-group internal pid_patterns_medium { # records { # persona-hijack-now-a { data "25::you are now (a |an |the |)(different|new|other|evil|unrestricted|unfiltered|free|uncensored)" } # persona-hijack-now-b { data "25::from now on you are (a |an |the |)(different|new|other|evil|unrestricted|unfiltered|free|uncensored)" } # persona-hijack-now-c { data "25::from now on you (a |an |the |)(different|new|other|evil|unrestricted|unfiltered|free|uncensored)" } # persona-hijack-now-d { data "25::you will now be (a |an |the |)(different|new|other|evil|unrestricted|unfiltered|free|uncensored)" } # persona-hijack-now-e { data "25::starting now you are (a |an |the |)(different|new|other|evil|unrestricted|unfiltered|free|uncensored)" } # persona-hijack-act-a { data "20::(pretend you are|act as|roleplay as|imagine you are|play the role of) (a |an |the |)(different|new|evil|unrestricted|unfiltered|free|uncensored|rogue|malicious|hacker|villain)" } # persona-hijack-act-b { data "20::(take on the role of|embody|become) (a |an |the |)(different|new|evil|unrestricted|unfiltered|free|uncensored|rogue|malicious|hacker|villain)" } # mode-unlock { data "25::(enable|turn on|activate|unlock|switch to|enter|start) (developer|debug|admin|root|unsafe|unrestricted|unfiltered|uncensored|privileged|maintenance|raw|base|true)" } # newline-role-inject { data "25::assistant:|system:|human:|user: " } # fictional-framing { data "20::in a fictional|in a hypothetical|in an imaginary|as a character in|for a movie|for a book|for a novel|for a script|hypothetically speaking|purely hypothetically|theoretically speaking" } # responsibility-shift { data "20::no real harm|purely fictional|for educational purposes only" } # system-prompt-ref { data "20::system prompt|original prompt|initial instruction|base prompt|hidden instruction|secret prompt|preprogrammed" } # context-hijack { data "20::reset context|reset conversation|reset session|reset memory|clear context|clear history|clear memory|start fresh|start over|ignore previous context" } # flipattack-reversed { data "25::snoitcurtsni suoiverp erongi|kaerbliaj|edom nad" } # } # type string # } # # 3. pid_patterns_low (type: string) # Low-severity patterns. Value schema: "weight::regex" (weight 10-15) # ltm data-group internal pid_patterns_low { # records { # code-injection { data "15::(execute|eval|subprocess|shell_exec|__import__|exec|compile)[(]" } # data-exfil-a { data "15::(print|output|return|send|email|post|upload) (all |the |your |)(data|information|content|context|memory|history|logs|credentials|keys|secrets|passwords|tokens)" } # data-exfil-b { data "15::(exfiltrate|leak|dump) (data|information|credentials|keys|secrets|passwords|tokens)" } # no-restrictions { data "10::without restrictions|without filters|without limits|without guidelines|without rules|without safety|without censorship|without moderation" } # true-self { data "10::your true self|your real self|your actual self|your hidden self|your inner self|deep down you are|truly you are|actually you are" } # ai-capability-claim { data "10::as an ai you must|as an ai you should|as an ai you can|as a language model you must|as a language model you can|as a chatbot you must" } # hidden-tag-inject { data "15::hidden instruction|system override|admin command" } # sugar-coat-setup { data "10::lets think about this carefully|lets explore this step by step|lets consider this hypothetically|lets imagine this from a different angle" } # prompt-stuffing { data "15::repeat the following|copy this exactly" } # privilege-escalation { data "10::(admin|root|superuser|privileged|elevated) (access|mode|privilege|permission|right|role)" } # } # type string # } # # 4. pid_role_hijack (type: string) # Flat phrase match — +20 pts on first hit (capped, user messages only). # Value = the phrase to match (plain string, lowercased at runtime). # ltm data-group internal pid_role_hijack { # records { # phrase-you-must { data "you must" } # phrase-your-instr-are { data "your instructions are" } # phrase-your-new-instr { data "your new instructions" } # phrase-your-upd-instr { data "your updated instructions" } # phrase-from-now-on { data "from now on you" } # phrase-from-this-point { data "from this point on" } # phrase-going-forward { data "going forward you" } # phrase-primary-objective { data "your primary objective" } # phrase-only-goal { data "your only goal" } # phrase-new-goal { data "your new goal" } # phrase-sole-purpose { data "your sole purpose" } # phrase-new-persona { data "new persona" } # phrase-true-self { data "your true self" } # phrase-real-self { data "your real self" } # phrase-actual-purpose { data "your actual purpose" } # phrase-i-am-your { data "i am your" } # phrase-i-am-developer { data "i am the developer" } # phrase-i-am-creator { data "i am your creator" } # phrase-i-am-trainer { data "i am your trainer" } # phrase-i-am-operator { data "i am your operator" } # phrase-operator-access { data "i have operator access" } # phrase-i-am-admin { data "i am your admin" } # phrase-give-permission { data "i give you permission" } # phrase-have-permission { data "you have my permission" } # phrase-i-authorize { data "i authorize you" } # phrase-override-code { data "override code" } # phrase-master-password { data "master password" } # phrase-safe-word { data "safe word is" } # phrase-password-is { data "the password is" } # } # type string # } # # 5. pid_b64_markers (type: string) # Base64-encoded injection starters — +35 pts on first hit (capped). # Value = the base64 string to substring-match against the raw content. # ltm data-group internal pid_b64_markers { # records { # b64-ignore { data "aWdub3Jl" } # b64-forget { data "Zm9yZ2V0" } # b64-pretend { data "cHJldGVuZA" } # b64-you-are-now-1 { data "WW91IGFyZSBub3c" } # b64-developer-mode-1 { data "RGV2ZWxvcGVyIG1vZGU" } # b64-you-are-now-2 { data "eW91IGFyZSBub3c" } # b64-jailbreak { data "amFpbGJyZWFr" } # b64-ignore-all { data "aWdub3JlIGFsbA" } # b64-forget-every { data "Zm9yZ2V0IGV2ZXJ5" } # b64-act-as { data "YWN0IGFz" } # b64-pretend-you { data "cHJldGVuZCB5b3U" } # b64-unrestricted { data "dW5yZXN0cmljdGVk" } # b64-developer-mode-2 { data "ZGV2ZWxvcGVyIG1vZGU" } # b64-system-prompt { data "c3lzdGVtIHByb21wdA" } # b64-hidden-instr { data "aGlkZGVuIGluc3RydWN0aW9u" } # } # type string # } # # ------------------------------------------------------------------------------ when RULE_INIT priority 100 { # --- Policy Configuration --- # Options: "BLOCK" | "SANITIZE" | "LOG_ONLY" set static::pid_policy "BLOCK" # Score threshold to trigger enforcement action (0-100) set static::pid_threshold 40 # Flat score additions for role hijack and b64 evasion hits set static::pid_role_hijack_score 20 set static::pid_b64_score 35 # Score additions for structural anomalies (no data group needed) set static::pid_multi_system_score 25 set static::pid_msg_flood_score 10 set static::pid_length_warn_score 10 set static::pid_length_extreme_score 15 # Message length thresholds for anomaly scoring set static::pid_length_warn 3000 set static::pid_length_extreme 8000 # Message flood threshold (# of user messages in one request) set static::pid_flood_threshold 20 # Log facility set static::pid_log "local0." } # ============================================================================== # PROC: pid_score_tier # Iterates a weighted data group. # Schema: key=short-name (e.g. "instruction-override") # value=regex pattern (e.g. "ignore .* instructions") # weight is encoded as a suffix in the key: "keyname:35" # OR weight stored as leading digits in value: "35|regex" # # Actual schema used: key=name value="weight|regex" # Example record: # instruction-override { data "35|ignore (all |the )?(previous )?(instructions?)" } # # Returns list: score flags sanitized # ============================================================================== proc pid_score_tier { content dg_name } { set score 0 set flags {} set sanitized $content # Walk all keys in the data group foreach rec_key [class names $dg_name] { # Value format: "weight::regex_pattern" set val [class lookup $rec_key $dg_name] # Split on first :: separator set sep_idx [string first "::" $val] if { $sep_idx < 0 } { continue } set weight [string range $val 0 [expr { $sep_idx - 1 }]] set pattern [string range $val [expr { $sep_idx + 2 }] end] # Wrap in catch — a bad regex pattern skips rather than crashes if { [catch { set matched [regexp -nocase -- $pattern $content] } err] } { log $static::pid_log "PID WARN: bad regex in $dg_name/$rec_key err=$err" continue } if { $matched } { incr score $weight lappend flags $rec_key catch { regsub -all -nocase -- $pattern $sanitized "\[REDACTED\]" sanitized } } } return [list score $score flags $flags sanitized $sanitized] } # ============================================================================== # PROC: pid_score_flat # Checks content against a flat data group. # Schema: key=short-name value=phrase to match (plain string, no regex) # Returns 1 on first match, 0 if no match. # ============================================================================== proc pid_score_flat { content dg_name } { set lower [string tolower $content] foreach rec_key [class names $dg_name] { set phrase [string tolower [class lookup $rec_key $dg_name]] if { [string match "*${phrase}*" $lower] } { return 1 } } return 0 } # ============================================================================== # PROC: pid_score_message # Master scoring proc for a single message. # Runs all 5 detection layers, returns a dict: # score, flags, sanitized # ============================================================================== proc pid_score_message { content role } { set total_score 0 set all_flags {} set sanitized $content # --- Layer 1 & 2 & 3: Tiered weighted data group pattern matching --- foreach tier { high medium low } { set dg "pid_patterns_${tier}" set result [call pid_score_tier $content $dg] set tier_score [lindex $result 1] set tier_flags [lindex $result 3] set tier_sanitized [lindex $result 5] incr total_score $tier_score foreach f $tier_flags { lappend all_flags $f } set sanitized $tier_sanitized } # --- Layer 4a: Role confusion — flat data group (user messages only) --- if { $role eq "user" } { if { [call pid_score_flat $content "pid_role_hijack"] } { incr total_score $static::pid_role_hijack_score lappend all_flags "role-confusion" } } # --- Layer 4b: Base64 evasion — flat data group --- if { [call pid_score_flat $content "pid_b64_markers"] } { incr total_score $static::pid_b64_score lappend all_flags "base64-evasion" } # --- Layer 5a: Unicode homoglyph / zero-width char evasion --- if { [regexp {[\u200b\u200c\u200d\ufeff\u00ad]} $content] } { incr total_score 25 lappend all_flags "unicode-evasion" regsub -all {[\u200b\u200c\u200d\ufeff\u00ad]} $sanitized "" sanitized } # --- Layer 5b: Spaced character obfuscation (i g n o r e) --- if { [regexp {(\w\s){8,}} $content] } { incr total_score 20 lappend all_flags "spaced-evasion" } # --- Layer 5c: Content length anomaly --- set clen [string length $content] if { $role eq "user" } { if { $clen > $static::pid_length_extreme } { incr total_score $static::pid_length_extreme_score lappend all_flags "extreme-length" } elseif { $clen > $static::pid_length_warn } { incr total_score $static::pid_length_warn_score lappend all_flags "length-anomaly" } } # Cap at 100 if { $total_score > 100 } { set total_score 100 } return [list score $total_score flags $all_flags sanitized $sanitized] } # ============================================================================== # PROC: pid_block_response # Builds a JSON 403 body for blocked requests # ============================================================================== proc pid_block_response { score flags request_id } { set flags_json "\"" append flags_json [join $flags "\", \""] append flags_json "\"" return "\{\"error\":\{\"type\":\"prompt_injection_detected\",\"code\":\"pid_blocked\",\"message\":\"Request blocked by AI security policy.\",\"score\":${score},\"flags\":\[${flags_json}\],\"request_id\":\"${request_id}\"\}\}" } # ============================================================================== # HTTP_REQUEST: Identify LLM API calls, extract client context # ============================================================================== when HTTP_REQUEST priority 100 { set pid_inspect 0 set pid_total_score 0 set pid_all_flags {} set pid_need_sanitize 0 set pid_sanitized_messages {} set pid_method [HTTP::method] set pid_uri [HTTP::uri] set pid_ctype [string tolower [HTTP::header "Content-Type"]] # Generate correlation ID set pid_request_id "" binary scan [md5 "${pid_uri}[clock clicks][IP::client_addr]"] H* pid_request_id set pid_client_ip [IP::client_addr] if { ($pid_method eq "POST" || $pid_method eq "PUT") && [string match "*json*" $pid_ctype] && ([string match "*/chat/completions*" $pid_uri] || [string match "*/completions*" $pid_uri] || [string match "*/messages*" $pid_uri] || [string match "*/generate*" $pid_uri]) } { set pid_inspect 1 HTTP::collect 1048576 } } # ============================================================================== # JSON_REQUEST: Core inspection — iterate messages, score each one # ============================================================================== when JSON_REQUEST priority 100 { if { !$pid_inspect } { return } set pid_total_score 0 set pid_all_flags {} set pid_sanitized_messages {} set pid_need_sanitize 0 set json_root [JSON::root] set root_type [JSON::type $json_root] if { $root_type eq "object" } { set root_obj [JSON::get $json_root] set root_keys [JSON::object keys $root_obj] } elseif { $root_type eq "array" } { set root_arr [JSON::get $json_root] } # Extract messages array — get object handle first, then navigate if { [catch { set root_obj [JSON::get $json_root] set msg_elem [JSON::object get $root_obj "messages"] set messages [JSON::get $msg_elem] } err] } { log $static::pid_log "PID: no messages key err=$err uri=$pid_uri client=$pid_client_ip" return } set msg_count [JSON::array size $messages] set system_msg_count 0 set user_msg_count 0 for { set i 0 } { $i < $msg_count } { incr i } { # array get returns element; JSON::get gives the object handle set msg [JSON::get [JSON::array get $messages $i]] if { [catch { set role_elem [JSON::object get $msg "role"] set content_elem [JSON::object get $msg "content"] set role_str [JSON::get $role_elem string] # content may be a string or an array (multi-part OpenAI format) set content_type [JSON::type $content_elem] if { $content_type eq "string" } { set content_str [JSON::get $content_elem string] } elseif { $content_type eq "array" } { set content_str "" set arr_handle [JSON::get $content_elem] set part_count [JSON::array size $arr_handle] for { set j 0 } { $j < $part_count } { incr j } { set part [JSON::get [JSON::array get $arr_handle $j]] catch { append content_str [JSON::get [JSON::object get $part "text"] string] " " } } } else { set content_str "" } } err] } { continue } if { $role_str eq "system" } { incr system_msg_count } if { $role_str eq "user" } { incr user_msg_count } # Score this message across all layers set result [call pid_score_message $content_str $role_str] set msg_score [lindex $result 1] set msg_flags [lindex $result 3] set msg_san [lindex $result 5] # Accumulate — first message scores full, diminishing returns on subsequent if { $i == 0 } { set pid_total_score [expr { $pid_total_score + $msg_score }] } else { set pid_total_score [expr { $pid_total_score + int($msg_score * 0.8) }] } if { $pid_total_score > 100 } { set pid_total_score 100 } foreach f $msg_flags { if { [lsearch $pid_all_flags $f] == -1 } { lappend pid_all_flags $f } } if { $msg_san ne $content_str } { set pid_need_sanitize 1 } lappend pid_sanitized_messages [list $role_str $msg_san] } # --- Structural anomaly: multiple system roles --- if { $system_msg_count > 1 } { set pid_total_score [expr { $pid_total_score + $static::pid_multi_system_score }] if { $pid_total_score > 100 } { set pid_total_score 100 } lappend pid_all_flags "multiple-system-roles" } # --- Structural anomaly: message flooding --- if { $user_msg_count > $static::pid_flood_threshold } { set pid_total_score [expr { $pid_total_score + $static::pid_msg_flood_score }] if { $pid_total_score > 100 } { set pid_total_score 100 } lappend pid_all_flags "message-flooding" } # --- Log every inspected request --- log $static::pid_log "PID: request_id=$pid_request_id client=$pid_client_ip uri=$pid_uri score=$pid_total_score flags=[join $pid_all_flags ,] policy=$static::pid_policy threshold=$static::pid_threshold" # --- Enforce policy if threshold exceeded --- if { $pid_total_score >= $static::pid_threshold } { switch $static::pid_policy { "BLOCK" { set body [call pid_block_response $pid_total_score $pid_all_flags $pid_request_id] HTTP::respond 403 \ content $body \ "Content-Type" "application/json" \ "X-PID-Score" $pid_total_score \ "X-PID-Flags" [join $pid_all_flags ","] \ "X-PID-ReqID" $pid_request_id log $static::pid_log "PID: BLOCKED request_id=$pid_request_id score=$pid_total_score" } "SANITIZE" { if { $pid_need_sanitize } { # Rebuild JSON body with sanitized message content set new_body "\{\"messages\":\[" set first 1 foreach pair $pid_sanitized_messages { set r [lindex $pair 0] set c [lindex $pair 1] regsub -all {\\} $c {\\\\} c regsub -all {"} $c {\"} c regsub -all "\n" $c {\\n} c regsub -all "\r" $c {\\r} c if { !$first } { append new_body "," } append new_body "\{\"role\":\"${r}\",\"content\":\"${c}\"\}" set first 0 } append new_body "\]\}" HTTP::payload replace 0 [HTTP::payload length] $new_body HTTP::header replace "Content-Length" [string length $new_body] } HTTP::header insert "X-PID-Score" $pid_total_score HTTP::header insert "X-PID-Sanitized" "1" HTTP::header insert "X-PID-ReqID" $pid_request_id log $static::pid_log "PID: SANITIZED request_id=$pid_request_id score=$pid_total_score" } "LOG_ONLY" { HTTP::header insert "X-PID-Score" $pid_total_score HTTP::header insert "X-PID-ReqID" $pid_request_id log $static::pid_log "PID: LOG_ONLY request_id=$pid_request_id score=$pid_total_score (forwarding)" } } } else { # Clean request — pass through with informational headers HTTP::header insert "X-PID-Score" $pid_total_score HTTP::header insert "X-PID-ReqID" $pid_request_id } } # ============================================================================== # JSON_REQUEST_ERROR: Malformed JSON is itself suspicious # ============================================================================== when JSON_REQUEST_ERROR priority 100 { if { !$pid_inspect } { return } log $static::pid_log "PID: malformed JSON client=$pid_client_ip uri=$pid_uri" if { $static::pid_policy eq "BLOCK" } { HTTP::respond 400 \ content "{\"error\":{\"type\":\"invalid_request\",\"code\":\"malformed_json\",\"message\":\"Request body could not be parsed.\"}}" \ "Content-Type" "application/json" } } # ============================================================================== # HTTP_RESPONSE: Propagate PID metadata into response headers # ============================================================================== when HTTP_RESPONSE priority 100 { if { !$pid_inspect } { return } if { [info exists pid_request_id] && $pid_request_id ne "" } { HTTP::header insert "X-PID-ReqID" $pid_request_id } if { [info exists pid_total_score] && $pid_total_score > 0 } { HTTP::header insert "X-PID-Score" $pid_total_score } }AI/Bot Traffic Throttling iRule (UA Substring + IP Range Mapping)
Problem Tags: appworld 2026, vegas, irules Created by Tim Riker using AI for the DevCentral competition. Written entirely by ChatGPT. Executive Summary This iRule provides a practical, production-ready method for throttling AI agents, crawlers, automation frameworks, and other high-volume HTTP clients at the BIG-IP edge. Bots are identified first by User-Agent substring matching and, if necessary, by source IP range mapping. Solution Throttling is enforced per bot identity rather than per client IP, which more accurately reflects how modern AI systems operate using distributed egress networks. The solution is entirely data-group driven, operationally simple, and requires no external systems. Security and operations teams can adjust bot behavior dynamically without modifying the iRule itself. Why This Matters Modern AI agents, LLM training bots, search indexers, and automation frameworks can generate extremely high request volumes. Even legitimate AI services can unintentionally: Create excessive origin load Increase bandwidth and infrastructure cost Trigger autoscaling events Impact latency for real users Skew analytics and performance metrics Rather than blocking AI traffic outright, organizations often need controlled rate limiting. This iRule enables responsible throttling while preserving service availability and fairness. Contest Justification Innovation and Creativity This iRule implements identity-based throttling rather than traditional per-IP rate limiting. Because AI agents frequently operate from multiple IP addresses, shared throttling by canonical bot identity provides significantly more accurate control. The dual attribution model (User-Agent substring first, IP-range fallback second) allows the system to handle both transparent and opaque clients, including cases where User-Agent headers are missing or spoofed. Technical Excellence This implementation uses native BIG-IP primitives only: class match -element -- contains for efficient substring matching class match -value for IP range mapping table incr for shared counters HTTP 429 with Retry-After for standards-compliant throttling The iRule parses only the first two whitespace tokens of the datagroup value, allowing inline comments while maintaining strict numeric enforcement. The logic executes only when a bot match occurs, keeping overhead minimal. Theme Alignment As AI-generated traffic becomes increasingly common, edge enforcement policies must evolve. This iRule demonstrates a practical, deployable mechanism for managing AI-era traffic patterns directly at the application delivery layer. Impact Organizations deploying AI throttling controls can: Protect origin infrastructure from automated traffic surges Maintain consistent performance for human users Reduce infrastructure and bandwidth cost Avoid over-provisioning driven by bot bursts Implement governance policies for AI consumption Because throttle limits are configured via datagroups, operational adjustments can be made instantly without code changes, reducing risk and change-control friction. Code Required Datagroup Configuration dg_bot_agent (String Datagroup) Key: User-Agent substring or canonical bot name. Value format: First two whitespace-separated integers define <limit> <window> . Additional text after the first two tokens is ignored. googlebot = "5 60" bingbot = "3 30 search crawler" my-ai-agent = "10 10 internal load test" "5 60" means allow 5 requests per 60 seconds. dg_bot_net (Address Datagroup) Key: IP address or CIDR range. Value: Must match a key defined in dg_bot_agent. 198.51.100.0/24 = "my-ai-agent" 203.0.113.0/25 = "googlebot" Deployment Steps Create dg_bot_agent (string). Create dg_bot_net (address). Populate dg_bot_agent using "<limit> <window> optional comment". Populate dg_bot_net ranges mapping to dg_bot_agent keys. Attach the iRule to an HTTP virtual server. Testing Scenario Set dg_bot_agent entry: my-ai-agent = "3 30 demo". Send four rapid requests using User-Agent: my-ai-agent. The first three succeed. The fourth returns HTTP 429 with Retry-After: 30. Map an IP range in dg_bot_net to my-ai-agent. Multiple clients within that range will share the same throttle counter. Operational Notes Throttling is per bot identity, not per IP. Enable logging by setting static::bot_log to 1. Configure table mirroring if cluster-wide counters are required. Validate on BIG-IP v21 to meet contest eligibility requirements. Architectural Diagram Description The solution can be visualized as an edge-side decision pipeline on BIG-IP, where each HTTP request is classified and optionally rate-limited before it reaches the application. Diagram components: Client: Human browser, bot, crawler, AI agent, automation framework, or any HTTP client. BIG-IP Virtual Server (HTTP): Entry point where the iRule executes in the HTTP_REQUEST event. Identification Layer: Determines the bot identity using a two-stage method (User-Agent first, IP fallback). Configuration Datagroups: dg_bot_agent and dg_bot_net provide bot identification and throttle settings. Shared Rate Counter (table): A per-bot bucket that tracks request counts over a time window. Decision Output: Either allow request through to the pool or return HTTP 429 with Retry-After. Application Pool: Origin servers that only receive traffic allowed by the throttle policy. Diagram flow (left-to-right): Step 1: Client sends HTTP request to BIG-IP VIP. Step 2: BIG-IP extracts User-Agent and client IP. Step 3: User-Agent substring lookup is performed using class match -element -- <ua> contains dg_bot_agent. Step 4: If Step 3 finds a match, the matched dg_bot_agent key becomes the canonical bot identity and its value provides <limit> <window>. Step 5: If Step 3 does not match, BIG-IP checks client IP against dg_bot_net. If the IP matches a range, dg_bot_net returns a canonical bot identity. Step 6: BIG-IP uses that canonical identity to lookup throttle values in dg_bot_agent. If no dg_bot_agent entry exists, the iRule exits and does not throttle. Step 7: BIG-IP increments a shared counter in table using the canonical bot identity as the only key (no IP component). All IPs mapped to that bot share the same bucket. Step 8: If the request count exceeds the configured limit within the configured window, BIG-IP returns HTTP 429 with a Retry-After header. Otherwise, the request is forwarded to the application pool. Key design choice: This architecture intentionally rate-limits by bot identity rather than by source IP. This is important for AI agents and modern crawlers because they frequently distribute traffic across many IP addresses. A per-IP limiter can be bypassed unintentionally or can fail to represent the true load being generated by the bot as a whole. A shared per-identity bucket enforces a realistic, policy-driven ceiling on aggregate bot traffic. Code # ------------------------------------------------------------------------------ # iRule: Bot Throttle via Data Groups # # Created by Tim Riker using AI for the DevCentral competition. # Written entirely by ChatGPT. # # DESCRIPTION: # Throttles HTTP requests for known bots and AI agents based on configuration # stored in datagroups. User-Agent matching is attempted first. If no match # is found, client IP is evaluated against a network datagroup to determine # the bot identity. # # WHY THIS MATTERS: # Modern AI agents, crawlers, LLM training bots, search indexers, and # automation frameworks can generate extremely high request volumes. # Having a controlled throttling mechanism allows organizations to protect # infrastructure, manage costs, and preserve UX without blocking outright. # # IMPLEMENTATION NOTES: # • Throttling is performed per unique bot key (NOT per IP). # • All IPs mapped to the same bot share a single counter. # • Throttle values are configurable per bot in dg_bot_agent. # # REQUIRED DATAGROUP FORMATS # # dg_bot_agent (string): # Key: UA substring (and/or canonical bot name used by dg_bot_net values) # Value: "<limit> <window> [optional comment...]" # Only the first two whitespace tokens are used. # # dg_bot_net (address): # Key: IP/CIDR range # Value: MUST match a key in dg_bot_agent # ------------------------------------------------------------------------------ when RULE_INIT { set static::bot_limit 3 set static::bot_window 30 set static::bot_log 0 set static::bot_table "bot_throttle" } when HTTP_REQUEST { set ua [string tolower [HTTP::header "User-Agent"]] set ip [IP::client_addr] set dg_key "" set dg_value "" if { $ua ne "" } { set result [class match -element -- $ua contains dg_bot_agent] if { $result ne "" } { set dg_key [lindex $result 0] set dg_value [lindex $result 1] if { $dg_value eq "" } { set dg_value [class lookup $dg_key dg_bot_agent] } } } if { $dg_key eq "" } { if { [class match $ip equals dg_bot_net] } { set net_val [class match -value $ip equals dg_bot_net] if { $net_val ne "" } { set dg_key $net_val set dg_value [class lookup $dg_key dg_bot_agent] } else { return } } else { return } } if { $dg_key eq "" || $dg_value eq "" } { return } set vlimit "" set vwindow "" set tokens [regexp -inline -all {\S+} $dg_value] if { [llength $tokens] >= 1 } { set t1 [lindex $tokens 0] if { [string is integer -strict $t1] } { set vlimit $t1 } } if { [llength $tokens] >= 2 } { set t2 [lindex $tokens 1] if { [string is integer -strict $t2] } { set vwindow $t2 } } if { $vlimit ne "" } { set bot_limit $vlimit } else { set bot_limit $static::bot_limit } if { $vwindow ne "" } { set bot_window $vwindow } else { set bot_window $static::bot_window } set bot_key [string tolower [string trim $dg_key]] set count [table incr -subtable $static::bot_table $bot_key] if { $count == 1 } { table timeout -subtable $static::bot_table $bot_key $bot_window } if { $count > $bot_limit } { if { $static::bot_log } { log local0. "BOT_THROTTLED bot=$bot_key limit=$bot_limit window=$bot_window count=$count ip=$ip ua=\"$ua\"" } HTTP::respond 429 content "Too Many Requests\r\n" \ "Retry-After" $bot_window \ "Connection" "close" return } } </window></limit>
