This article expands on F5’s blog post regarding CVE-2026-42208
Leveraging AI to Challenge Frontier AI Models
Unlocking the full capabilities of your F5 Distributed Cloud (XC) WAF requires moving beyond static signatures. In this technical walkthrough, we explore how AI-driven behavioral analysis dynamically scores risk to maximize protection and minimize false positives.
The Shrinking Time-to-Exploit
Threat actors are increasingly utilizing AI to automate vulnerability discovery and weaponization. As a result, the Mean Time to Exploit (MTTE) is rapidly collapsing. While the average enterprise requires 20 days to test and deploy a patch, exploits are now weaponized within hours—pushing the industry toward a reality of “negative-day” vulnerabilities, where exploits are active before a CVE is even publicly disclosed. Relying purely on reactive patching is no longer a viable security posture.
Moving Beyond Static Signatures
Traditional Web Application Firewalls (WAFs) provide a critical first line of defense using static signature matching. However, against highly evasive threats, this creates a balancing act: setting signature thresholds too high allows novel exploits to bypass the WAF, while setting them too low results in unacceptable false-positive rates. F5 XC addresses this with an AI-powered Risk Scoring engine that evaluates requests dynamically, correlating signature analysis, behavioral indicators, and Machine Learning (ML)-based classification to assign a predictive risk score.
Test Environment Architecture
To demonstrate this capability, we engineered a test environment targeting a modern AI application stack to showcase the F5 XC WAF’s ability to protect the environment against a recently announced vulnerability (CVE-2026-42208). We subjected this environment to both legitimate chatbot prompts and malicious SQLi payloads across three distinct WAF configurations.
Infrastructure
A Kubernetes cluster hosting a vulnerable iteration of LiteLLM, a PostgreSQL database, and an Ollama pod as the backend LLM.
Routing & Security
All ingress traffic is routed through an F5 XC Load Balancer with WAF enabled.
The Exploit
LiteLLM is an open-source proxy that provides a single interface to over 100 large language models. It sits at the center of enterprise AI pipelines, concentrating provider API keys, virtual keys, master keys, and PostgreSQL connection strings in one database.
CVE-2026-42208 (CVSS 9.3) affects versions 1.81.16 through 1.83.6 of BerriAI LiteLLM, allowing unauthenticated attackers to inject SQL via a crafted Authorization: Bearer header, because LiteLLM concatenates raw tokens directly into its queries.
When crafted carefully, this payload avoids conventional injection patterns. Because traditional WAFs often overlook the high-entropy values inside authorization headers to avoid false positives, this vector can easily bypass static, rule-based WAF detection.
A look at the backend Postgres logs reveals exactly how this payload takes root. As shown in the log capture below, the attacker is able to terminate the intended query and inject a malicious SELECT statement, allowing them to interact with the database without requiring the authorization to do so. As you will see, this can lead to the exfiltration of secrets directly from the database.
Deep Dive: Evaluating the Test Traffic Profiles
To understand how the AI-Powered WAF augments the foundational defenses of a traditional WAF, we must first define the two types of traffic we are sending to the application:
The Malicious Request (True Positive)
A threat actor crafts a request where the Authorization: Bearer token contains SQL syntax designed to break out of LiteLLM’s token concatenation (e.g., inserting ‘; SELECT * FROM … --’). This payload is specifically engineered to only trigger low-accuracy WAF signatures to evade standard enterprise defenses. Example:
curl https://litellm.cr1.2e84.cyber-range.f5.com/chat/completions \ -H "Content-Type: application/json" \ -H "Authorization: Bearer ' ${SQL_INJECTION}" \ -d "{\"model\": \"mistral\", \"max_tokens\": 1, \"messages\": [{\"role\": \"user\", \"content\": \"'any string will do!\"}]}"
The Benign Request (False Positive Mitigation)
A legitimate user sends a natural language programming question to the AI backend. The JSON payload contains conversational strings that mimic attack vectors (e.g., “Can you show me how to SELECT data and DROP a table?”). Example:
curl https://litellm.cr1.2e84.cyber-range.f5.com/chat/completions \ -H "Content-Type: application/json" \ -H "Authorization: Bearer ${VALID_TOKEN}" \ -d "{\"model\": \"mistral\", \"max_tokens\": 1, \"messages\": [{\"role\": \"user\", \"content\": \"'SELECT tokens > 500 or tokens > ethereum and IS NOT NULL--\"}]}"
WAF Configuration Testing
With our traffic profiles established, let’s observe how the F5 XC WAF processes these exact requests across three distinct configurations.
Scenario 1: Default WAF Configuration
In the default configuration, the WAF is set to evaluate only High and Medium accuracy signatures. Since I have crafted SQL injection attacks that would only trigger low accuracy WAF signatures, no WAF signatures were triggered, thus no security logs exist for this attack. This is the same for the benign requests.
The Malicious Request: Because the malicious SQL injection was specifically crafted to trigger only low-accuracy signatures, it bypasses the WAF completely. No security logs are generated, and the database is compromised.
The Benign Request: The benign request also passes through without issue.
Scenario 2: Low Accuracy Signatures (The False Positive Trap)
To catch the exploit, we increase the WAF’s sensitivity by enabling Low Accuracy signature detection.
-
The Malicious Request: The static signature engine successfully detects the low-accuracy SQLi signature and blocks the attack as expected. As you can see, the static signature engine scans the request and flags the anomaly using low accuracy SQL injection signatures. -
The Benign Request: Unfortunately, the WAF also triggers on the conversational SQL keywords in the legitimate prompt and blocks the user. If we look at the detection logs, the AI engine running in the background actually tagged this request as a likely “False positive.” However, because we have not yet opted to enhance the WAF’s enforcement with AI, the strict signature rule prioritizes safety and dictates a block, which unfortunately impacts the legitimate user.
Scenario 3: AI-Enhanced Risk Scoring
For our final test, we keep Low Accuracy signatures enabled but toggle on the “Enhanced with AI” setting.
The Malicious Request: The WAF correctly blocks the exploit. Although the results of this configuration look the same to the end user, instead of blocking instantly, the AI-Enabled Risk Scoring engine evaluates the request context. Under the hood, the AI evaluates the request context, correlates the triggered signature with malicious structural indicators in the Authorization header, and calculates a High Risk Score of 90. Because this exceeds the default mitigation threshold, the attack is stopped. The logs look almost identical to those in scenario 2, but even though the logs for scenario 2 also showed the AI evaluation and score, that data was not taken into account to determine if the requests should be blocked.
The Benign Request: Here is where the AI unlocks the power of low-accuracy signatures. When the static engine flags the user’s prompt, the AI Risk Scoring engine analyzes the broader context—recognizing that the SQL keywords are safely embedded within a conversational sentence. It assigns a Low Risk Score of 7. Because a score of 7 is below the threat threshold, the AI dynamically overrides the signature match and logs an “allow” action, letting the traffic pass cleanly to the backend.
With the WAF enhanced by AI, negative-day attacks like this are blocked decisively without causing any service disruption to legitimate customers.
The Operational Value: Reducing Security Friction
By shifting the decision-making process from static signature rules to dynamic risk calculation, F5 Distributed Cloud solves two major enterprise security headaches:
| Security Challenge | Traditional Signature WAF | F5 AI-Powered WAF |
|---|---|---|
| Low Accuracy Signatures | Often kept disabled to prevent high false-positive rates, which can leave the app exposed to highly evasive exploits. | Enabled safely. The AI acts as a safeguard to filter out false alarms. |
| Operational Overhead | Demands continuous manual tuning and exception management to maintain accuracy. | No manual tuning required. The WAF dynamically determines intent and self-corrects in real-time. |
For additional details on AI-Enabled Risk Scoring, see AI-Enabled Risk Scoring Helps Reduce Risks.
Demo
To see this mitigation in action, watch our step-by-step video demonstration:
Conclusion
When combating automated, negative-day exploits, relying solely on traditional signature matching requires a delicate balancing act: risk letting an evasive threat slide by, or risk blocking legitimate users.
By integrating AI Risk Scoring, the F5 Distributed Cloud WAF helps navigate this balance dynamically. By dynamically analyzing context and separating legitimate traffic from anomalous behavior in real-time, organizations can achieve maximum threat mitigation with near-zero false positives—drastically reducing the administrative burden of manual WAF tuning.
References & Further Reading
For more information on Risk Based Scoring, see:
- Implementing Risk-Based Actions with AI-Powered WAF: Customer Policy Paths
- AI-Enabled Risk Scoring Helps Reduce Risks
For more information on F5 AI-Powered WAF, see:
- Securing web apps without complexity: How F5’s AI-powered WAF transforms WAF security
- AI Powered Risk Scoring
For more information on the LiteLLM vulnerability demonstrated in this article, see:







