Attackers are finding vulnerabilities at machine speed with frontier AI models. Security teams are still pasting reports into chat windows and asking LLMs to "find the bugs."
Cisco is open-sourcing a fix for that gap. The Foundry Security Spec, published Tuesday by Cisco distinguished engineer Omar Santos, is a model-agnostic blueprint for building agentic security evaluation systems. It wraps frontier LLMs like Anthropic's Mythos and OpenAI's GPT-5.5-Cyber in orchestration, defined roles, and guardrails so detection and validation happen on purpose, not by accident in a chat window.
"One is an interesting demo," Santos wrote. "The other is a security evaluation system you can defend in front of your CISO and your auditors." The spec arrives as a pair of artifacts on GitHub, designed to work with GitHub's spec-kit development workflows. The "spec" artifact defines eight core agent roles (Orchestrator, Indexer, Cartographer, Detector, Triager, Validator, Coverage-Guide, Reporter), five extension roles, and roughly 130 functional requirements, each with an inline rationale. The "constitution" artifact lays out 11 inviolable principles, each encoding a real production failure Cisco shipped, diagnosed, and fixed.
Cisco isn't selling a tool. It's publishing a design. The company's internal implementations are too tied to its own infrastructure to open source directly, so the spec transfers the architecture: which roles are needed, what each must guarantee, how findings flow from detection to publication, and where quality gates go.
Cisco chief security and trust officer Anthony Greco said the company's access to Anthropic's Mythos showed a "material step forward" in model capabilities. Cisco is one of a small group of entities with Mythos access.
"We feel a lot of privilege to be in the position that we are, but we also feel a lot of responsibility," Cisco chief product officer Jeetu Patel told SDxCentral. "We're working round the clock." The spec pairs with Project CodeGuard, which Cisco previously open-sourced and donated to the Coalition for Secure AI (CoSAI). CodeGuard provides security rules for AI coding agents and feeds into Foundry's Detector role. The combined system creates a detection-to-prevention loop: Foundry's exploratory agents hunt for what no rule describes, CodeGuard catches that class of bugs on the next sweep, and the same ruleset loads into developers' editors to prevent the bug before it's ever written.
Greco framed the release bluntly in a prerecorded video. "Cybersecurity is a team sport," he said.
"This is one really demonstrable way where we're trying to raise the bar for everybody and share our knowledge through this."
Foundry Security Spec lives on GitHub under CiscoDevNet/foundry. It is not a turnkey scanner.
It is a seed and a starting point, meant to be adapted to each organization's environment and threat model.
