Cisco Open-Sources Agentic AI Security Spec

Why It Matters

By standardizing AI security evaluation, the spec helps enterprises turn fast‑moving LLM insights into reliable, auditable defenses, reducing false positives and compliance risk. It also accelerates industry adoption of safe AI practices, a growing priority as generative models proliferate in security operations.

Key Takeaways

• Cisco releases Foundry Security Spec as open-source on GitHub.
• Spec defines eight core agent roles and 130 functional requirements.
• Framework is model‑agnostic, enabling consistent AI security evaluation across LLMs.
• Includes auditable provenance chain and guardrails to prevent hallucinated findings.
• Works with Cisco’s CodeGuard to secure AI‑generated code throughout development.

Pulse Analysis

The rapid adoption of large language models (LLMs) in security operations has outpaced the development of robust evaluation processes, leaving teams vulnerable to noisy outputs and hallucinations. Cisco’s decision to open‑source the Foundry Security Spec addresses this gap by providing a structured, repeatable methodology that integrates orchestration, role definition, and safety constraints directly into AI workflows. By publishing the spec on GitHub and aligning it with the industry‑wide spec‑kit, Cisco invites collaboration and rapid iteration, positioning the framework as a de‑facto standard for AI‑driven threat detection.

At the heart of the spec are eight core agent roles—such as orchestrator, detector, and validator—supported by roughly 130 functional requirements and a constitution of eleven real‑world failure principles. This granular design ensures that every finding is bounded, prioritized, and verifiable, while an auditable provenance chain tracks detection through triage to publication. Because the framework is deliberately model‑agnostic, organizations can apply it to today’s frontier LLMs and future reasoning agents without re‑engineering the underlying guardrails, preserving consistency and compliance across evolving AI capabilities.

Cisco pairs the Foundry Security Spec with its complementary open‑source CodeGuard project, extending protection from AI‑generated code to the entire development lifecycle. CodeGuard injects secure‑by‑default rules into design, generation, and review phases, enabling tools like GitHub Copilot or Claude Code to automatically enforce best‑practice patterns. Together, these initiatives signal a maturing ecosystem where AI security is treated as a shared responsibility, offering enterprises a tangible path to trustworthy, auditable AI deployments while fostering industry‑wide collaboration.