Technical Articles
F5 SMEs share good practice.
Showing results for 
Search instead for 
Did you mean: 


Within the wide-reaching realm of machine learning, generative AI has recently taken a significant and cutting-edge role. Systems such as ChatGPT, Google Bard, and Microsoft Bing have quickly become technical advisors for content creators, system architects, and software developers. The use of Large Language Models (LLM) empowers algorithms to develop application source code or craft intricate narratives, thereby revolutionizing numerous digital business processes. However, the disruptive potential of generative AI also warrants careful consideration of security measures to protect against potential threats.

This article aims to highlight the significance of securing generative AI technology by assessing its assets, identifying potential threats, and recommending mitigation strategies. It aims to provide valuable insights and guidance on protecting the Confidentiality, Integrity, and Availability of generative AI in various applications.

Assets, Threats, and Mitigations

Assets, threats, and mitigations are essential components of threat modeling, protecting valuable assets and ensuring strong system security. Simply put, the security of our assets will face many threats, so we look to apply mitigations to protect our valuable data and systems. This applies generally to information security and generative AI systems are no exception.

  • Assets refer to the resources, systems, or data that need protection from threats
  • Threats are the potential risks or attacks that can compromise the security of these assets
  • Mitigations are the countermeasures put in place to minimize or eliminate the risks posed by threats

Inside Generative AI: Understanding Key Assets

Generative AI encompasses several key assets that contribute to its functionality and effectiveness. These assets include the infrastructure required to deliver services, the training data, the AI models, and well as the outcomes or outputs generated by the models. It is crucial to recognize and understand these assets within generative AI, as they have significant implications for system security.

Infrastructure: Infrastructure forms a pivotal asset in generative AI, encompassing the hardware like servers, GPUs for computation, and the software frameworks for model development, training, and deployment. It also includes the digital interfaces delivering AI-generated outputs to end-users. A reliable, high-performing, and secure infrastructure is key to enabling sophisticated generative AI models to operate effectively and provide value.

Training Data: High-quality training data is essential for training generative AI models. The training data acts as a source of knowledge and inspiration for the models to learn from. It typically consists of a diverse and representative set of examples that the model can use to understand the underlying patterns, styles, or characteristics it should capture. The quality, quantity, and diversity of the training data play a significant role in shaping the capabilities and generalization abilities of generative AI models.

AI Models: The AI models serve as a fundamental component and essential asset in generative AI, serving as the foundation for generating the desired outputs and enabling the technology to function effectively. These models, such as generative adversarial networks (GANs) or Transformer based models, are designed to learn patterns and relationships within the training data and generate new outputs based on that understanding. The architecture, parameters, and structure of these models are fundamental components that enable the generation of novel content.

Generated Outcomes: Generative AI outputs are valuable assets for businesses. When coupled with human experts, the generated outcomes have the ability to drive creativity, content generation, and data augmentation to influence strategic decision-making. The uniqueness and quality of these outcomes can give businesses a competitive edge, inspiring innovation and unlocking new opportunities in a dynamic marketplace.


Fortifying Foundations: Exploring the Role of Infrastructure in Generative AI

The potential challenges that could endanger this vital infrastructure are numerous and this section quite frankly won't try to cover the entire spectrum of threats. There are many other great bodies of work that cover threats to your infrastructure so I'll highlight a few for this key asset before moving into the AI specific threats.

Service Disruption

Threat: A significant threat to consider is "Denial of Service," where hardware malfunctions, software glitches, or network interruptions can markedly impact the operation of generative AI models. Such disturbances can result in service unavailability, potential loss of vital data, and a compromised ability of the model to learn, generate outputs, or interface with other systems. The repercussions can be considerable, particularly in applications demanding constant uptime or real-time processing.

Mitigation: To counteract potential Denial of Service disruptions, it is crucial to build redundancy into the system. This can involve having backup servers and fail-safe protocols to ensure persistent availability. Regularly updating software components and hardware devices can also help to avert potential vulnerabilities. Additionally, constant monitoring of system performance and capacity can enable early detection and swift resolution of issues, thereby minimizing service downtime and preserving the robustness of the AI infrastructure. 

Unauthorized Access

Threat: Intrusions into the system infrastructure may lead to malicious activities such as data theft, service disruption, or malicious code insertion. This not only risks the security of the AI models and data but can also result in the generation and spread of inaccurate or harmful outputs.

Mitigation: To fend off unauthorized access, a multi-faceted security approach is crucial. This should involve robust authentication protocols, proactive vulnerability management including regular software updates, and continous monitoring for early detection and prevention of intrusion attempts. A well-formulated incident response strategy is also vital, ensuring immediate action to limit the impact of any breaches and expedite system recovery.


Enhancing Data Resilience: Mitigating Threats to Training Data

The quality and security of training data are critical considerations as threats to training data can have profound implications for the performance and trustworthiness of generative AI models.

Data Quality and Bias

Threat: The quality and bias of training data directly impact on generative AI models, including risks associated with Training Data Poisoning along with causing issues where there is an Overreliance on LLM-generated Content. Poor quality data and biases in the training data can hinder the model's ability to learn accurate representations and produce reliable outcomes.

Mitigation: Addressing data quality and bias requires rigorous preprocessing, such as data cleaning, normalization, and augmentation. Techniques for bias detection and mitigation can also help reduce biases. Implementation of robust error handling mechanisms can help mitigate errors and data poisoning. Essential to this process is a 'human-in-the-loop' approach, which provides an extra layer of monitoring and adjustment, ensuring higher quality and bias control.

Intellectual Property Infringement

Threat: The unauthorized use or improper sourcing of training data can lead to intellectual property infringement, violating copyright or intellectual property rights. This exposes organizations to legal consequences, reputational risks, and loss of confidential data.

Mitigation: Implementing clear data usage policies, obtaining proper rights and permissions for the training data, and conducting thorough due diligence to ensure compliance with copyright and intellectual property laws are crucial steps to mitigate the risks of intellectual property infringement and protect the legal interests of all stakeholders involved.

Data Breaches and Privacy Concerns

Threat: As with other business critical data, the storage and handling of AI model training data have risks of data breaches, including Data Leakage, where unauthorized access or malicious attacks can compromise the security of sensitive information. 

Mitigation: Countering these risks necessitates robust data security measures. Encryption techniques and stringent access controls help protect data, while regular security audits identify potential vulnerabilities for swift resolution. Advanced methods like differential privacy and federated learning add extra layers of protection, maintaining privacy without hindering AI training.


Building a Fortress: Protecting the Heart of Generative AI Models

AI Models are not immune to threats and face potential risks that can undermine their integrity and reliability, jeopardizing the effectiveness and trustworthiness of generative AI technology. 

Adversarial Attacks

Threat: Adversarial attacks such as Prompt InjectionsServer Side Request Forgery (SSRF) and Unauthorized Code Execution pose significant threats to AI models in generative AI.

  • Prompt Injections allow malicious actors to manipulate the model's inputs by injecting carefully crafted prompts that make the model ignore previous instructions or perform unintended actions.
  • Server Side Request Forgery allows attackers to perform unintended requests or access restricted resources, possibly allowing access to "internal only" system interfaces.
  • Unauthorized Code Execution, as the name implies, involves exploiting the model to execute malicious code or actions on the underlying system.

Mitigation: To effectively mitigate the threats of prompt injections, SSRF vulnerabilities, and unauthorized code execution, a multi-layered defense approach is essential. Operators should implement a combination of specific techniques and security measures to ensure robust protection.

  • To prevent prompt injections, techniques such as prompt sanitization, input validation, and prompt filtering ensure that the model is not manipulated by maliciously crafted prompts, safeguarding the integrity of the generated outcomes.
  • For mitigating the risks associated with SSRF vulnerabilities, carefully validating and sanitizing incoming requests and strong network security measures, including network isolation and proper firewall configurations which restrict outbound requests play a crucial role in preventing SSRF attacks.
  • Risks of unauthorized code execution can be reduced by employing secure coding practices, conducting thorough code reviews, and utilizing runtime defenses like code sandboxing. These measures ensure the AI model runs on secure code and restrict unauthorized actions, providing enhanced system protection.

Model Theft or Replication

Threat: The unauthorized duplication or theft of AI models constitutes a significant threat. This can occur when there is unauthorized access to the model's parameters, architecture, or training data, potentially undermining its intellectual property and competitive edge.

Mitigation: A combination of robust access controls, encryption methods, and secure storage can help protect against model theft or replication. Additionally, techniques like watermarking or digital rights management can further safeguard the model's intellectual property. Regular monitoring and audits play a crucial role in promptly detecting and responding to unauthorized access attempts.


Promoting Dependable Results: Enhancing the Resilience of Generated Outcomes

The generated outcomes produced by generative AI models can considerably influence a wide array of sectors and industries, yet they are also susceptible to numerous threats.

Agent Manipulation

Threat: Manipulation of retrieval augmented generation (RAG) models or applications built on frameworks like Langchain, presents complex risks to application integrity and reliability. This threat involves the tampering with one or more of the facets involed in the information retrieval process, the injection of biased or misleading information or in some cases the intentional execution of code returned by LLMs.

Mitigation: To counteract agent manipulation, a layered defense approach is necessary. This includes implementing robust access controls, audit mechanisms, and employing ephemeral systems for isolation and resource management. These measures prevent unauthorized manipulations, ensure system accountability, and effectively contain potential spread of malicious code.

Compromised Model Supply Chain

Threat: Just like we see in the software world, the use of pretrained models in AI systems introduces many potential risks in the supply chain. One of the key threats is the possibility of compromised or malicious models being incorporated into AI systems. These pretrained models, developed and distributed by various organizations, may contain vulnerabilities or intentional backdoors that can lead to unintended consequences and compromise the integrity of the system. Such risks include biased or unreliable outputs, privacy breaches, and even the execution of unauthorized code. These threats can have far-reaching implications, affecting the trustworthiness and functionality of AI systems that rely on pretrained models.

Mitigation: Mitigating risks in pretrained model supply chains involves rigorous vetting, robust security measures, transparency, continuous monitoring, and collaboration within the AI community. To mitigate these risks, organizations should implement stringent vetting processes to select models from trusted sources, conducting thorough due diligence to assess security practices. Robust security measures should be in place, including secure transmission and storage of models, strong access controls, and regular security audits. Transparency should be promoted through model documentation, code review, and independent audits, allowing stakeholders to make informed decisions. I believe this is an open problem that can only be solved through with the AI community sharing knowledge and developing standards to ensure the integrity and ethical use of pretrained models.

Misinformation and Fake Content

Threat: The generation of outcomes through generative AI models introduces the risk of creating convincing fake content, also called "hallucinations" as well as the potential for generating outcomes that contain Prompt Injections. I personally believe confabulation is the more accurate description of this behavior but I use the more commonly accepted term of hallucinations. This can be exploited by malicious actors to deceive and manipulate the public, posing significant risks to public trust, reputation, and the integrity of information sources. As more of the Internet content becomes generated by AI systems, the more it becomes a feedback loop where AI generated content is training tomorrows models with AI generated content.

Mitigation: Robust content verification mechanisms, fact-checking processes, and responsible dissemination practices, including addressing prompt injections, are crucial in combating the spread of misinformation and fake content generated by AI models. I consider this an open problem that needs more research and collaboration amongst the community.



In the ever-progressing landscape of AI technology, generative AI has emerged as a powerful force, revolutionizing industries with its ability to generate unique outputs and drive innovation. However, this disruptive potential also brings forth security challenges that demand our attention. As we journey deeper into the current AI era, it is vital to remain vigilant, proactively staying ahead of potential threats and building resilient systems. Through persistent research and development in AI security, we can foster a future where generative AI is fully harnessed, unleashing its benefits in hybrid IT environments. By understanding and safeguarding our assets, identifying and mitigating potential threats, and upholding ethical practices, we can pave the way for secure and trustworthy deployment of generative AI. Together, let us embrace this transformative technology with unwavering dedication to security, forging a path towards a future where generative AI flourishes and positively shapes our world.


F5 Employee
F5 Employee

Thank you Jordan for this informative and helpful article!  You know that I have a healthy suspicion for fads in just about any sphere, but with all the hype around LLM's, I worry that we are about to plumb the depths of the possible negative outcomes of Overreliance on LLM-generated Content and Inadequate AI Alignment.  We're already seeing the negative outcomes of other AI based technology, like facial recognition systems that fail to identify people of color, autonomous vehicles that fail to see pedestrians, and that lovely fictional "image enhance" feature being realized, but with a serious racial bias.  I worry that humans with the latest new toy are going to (as we always do) apply it in increasingly inappropriate use-cases and cause widespread harm.  I'd like to encourage anyone to consider the following questions when applying LLM's or really any of the non-linear data processing (or AI) technologies:

1.  Does my use-case require that the outcomes *always* be correct (e.g. autonomous cars or kill-bots) or is some error acceptable?

2. If my new AI based system is sometimes wrong, will that cause more harm than the good it can do?

3. How much error is acceptable in my use-case?

4. Is there some way I can prevent my new AI based system from being wrong in a way that is completely convincing and thus causing exceptional harm?


Community Manager
Community Manager

Really well laid out. Great article @Jordan_Zebor !!

Version history
Last update:
‎30-Jun-2023 20:01
Updated by: