Cisco's Foundry Security Harness Specification Released

• •Cisco released the Foundry security harness specification for evaluating LLMs.
• •The specification allows for organization-neutral implementation and governance.
• •A proof of concept demonstrates the specification against real-world targets.

Cisco's Advanced Security Initiatives Group has released the Foundry security harness specification, which serves as a framework for evaluating large language models (LLMs) against security standards. The specification is designed to be organization-neutral and allows users to build their own implementations based on the architecture and guardrails provided. The Foundry system utilizes CodeGuard rules for detection, which Cisco has previously open-sourced. A proof of concept repository has also been created by a user, demonstrating the specification against several real targets, including Langflow and MLflow. This implementation aims to enhance governance and repeatability in security processes, producing verifiable and auditable outputs from LLMs. The Foundry specification is intended to help organizations improve their security posture without altering the underlying models.