F5 AI Security in Action - Part 2: F5 AI Remediate and F5 AI Guardrails

Key Concepts

• Inference-based Guardrails: A scalable security approach that applies policies at the inference layer rather than modifying the underlying RAG (Retrieval-Augmented Generation) or engineering pipeline.
• Purple Team Agents: Autonomous AI agents that simulate both offensive (red) and defensive (blue) tactics to identify vulnerabilities and generate optimized guardrails.
• Human-in-the-Loop (HITL): A verification process where security teams review and tweak AI-generated guardrails before production deployment.
• Project-based Segmentation: Organizing guardrails into specific "projects" to apply context-aware security policies (e.g., HR vs. public-facing assistants).
• Natural Language Guardrails: Defining security policies using plain language descriptions rather than relying solely on rigid regex or keyword matching.

1. Remediation Strategy and Philosophy

F5 emphasizes an inference-first approach to security. Instead of forcing engineering teams to rebuild RAG architectures or data enrichment processes every time a breach is identified, the security team deploys "guardrails" directly into the inference layer. This significantly reduces the "time to glass" (the time taken to secure the system) and allows security teams to harden applications without disrupting the core development workflow.

2. The Purple Team Agent Workflow

The remediation process is automated through the use of Purple Team Agents:

1. Identification: The system identifies a vulnerability (e.g., unauthorized access to salary data).
2. Generation: The agent analyzes the attack data and generates multiple variations of potential guardrails.
3. Testing: The agent tests these variations against the vulnerability to determine the highest prevention rate and accuracy.
4. Selection: The agent presents the most effective guardrail(s) to the human operator.
5. Refinement: The human operator reviews the suggestions, performs minor tweaks, and verifies the logic before deployment.

3. Technical Implementation and Testing

• Natural Language Definition: Guardrails are defined by describing the intent (e.g., "mentions of compensation terms such as salary, wages, tips, and bonuses").
• Iterative Tuning: The demonstration highlighted that guardrails must be generalized to be effective. A narrow guardrail (e.g., blocking only "Alice Johnson's salary") is easily bypassed by variations like "salary ranges" or "annual income." Effective guardrails require iterative testing and versioning to ensure they catch semantic variations of the prohibited intent.
• Versioning: Enterprise customers utilize versioning to benchmark and improve guardrail performance over time, ensuring that updates do not break existing workflows.

4. Project-Based Segmentation

Guardrails are managed within "Projects," which allow for context-specific security:

• Contextual Isolation: An HR assistant requires different guardrails than a public-facing makeup assistant.
• Regulatory Compliance: Projects can be segmented by geography (e.g., EU vs. US) to comply with regional regulations, such as upcoming EU AI legislation.
• Role-Based Access Control (RBAC): Projects can be tied to user roles. For example, an HR admin might have different access permissions than a standard user, and the guardrails can be configured to reflect these authorization levels.

5. Key Quotes

• "We don't want to go back to the engineering team and say, 'Right, you have breaches. You have to go and re-do your RAG, your enrichment.'... We want to put a guardrail in place. We can do that without having to factor into the [engineering] sprints."
• "The purple team has done 90% of the heavy lifting. The human just has to do maybe a bit of tweaking and verify on their side before they push it into production."

6. Synthesis and Conclusion

The F5 approach shifts the security paradigm from reactive engineering changes to proactive, inference-level policy enforcement. By leveraging Purple Team Agents, organizations can automate the discovery and creation of security guardrails, significantly reducing the burden on engineering teams. The combination of natural language policy definition, iterative testing, and project-based segmentation provides a robust framework for securing LLM-based applications while maintaining the necessary human oversight to ensure accuracy and compliance.