Cisco has released the open-source Foundry Security Spec to help organisations build verifiable AI-powered cyber defense systems with structured guardrails, auditable findings, and cross-LLM compatibility.
Cisco has open-sourced the Foundry Security Spec, an agentic AI cybersecurity evaluation framework designed to generate verifiable, auditable, and structured security findings against machine-speed cyber threats and vulnerabilities introduced by frontier AI models.
Available on GitHub, the community-driven framework is positioned as part of a broader push toward open AI security standardisation and collaborative cyber defense. The specification is model-agnostic and stack-neutral, enabling compatibility across different large language models and infrastructure environments, including systems such as Anthropic Mythos and OpenAI GPT-5.5-Cyber.
The release includes two core components: Spec.md, which defines eight core agent roles, five extension roles, and nearly 130 functional requirements; and Constitution.md, which outlines 11 inviolable principles derived from real-world failures encountered by Cisco.
The framework aims to address limitations in traditional “find and patch” security workflows, which often struggle against AI-driven attacks and hallucinated outputs from frontier LLMs. Cisco said Foundry wraps AI models with structured orchestration, governance layers, and guardrails to produce bounded, prioritised, and verifiable findings with clear completion signals.
Foundry also integrates with GitHub’s spec-kit and can pair with Project CodeGuard to create a continuous detection-to-prevention cycle, enabling organisations to convert discovered vulnerabilities into reusable protections against future bug classes.
Cisco noted that Foundry is not a turnkey scanner and still requires organisation-specific implementation and human oversight.
