--- title: README sdk: static pinned: false license: mit ---

Cognium Labs

# Cognium Labs **Trust infrastructure for code built by AI agents.** Cognium audits, verifies, and enforces AI-generated code before it reaches production. [Website](https://cognium.net) · [Open-source SAST](https://cognium.dev) · [Benchmarks](https://cognium.dev/benchmark) · [GitHub](https://github.com/cogniumhq/circle-ir)
## What we build Cognium helps engineering, security, and platform teams answer one release-critical question: **should this AI-generated change ship?** Our work combines: - Semantic SAST that traces real source-to-sink data flows instead of matching text patterns. - LLM-assisted verification for exploitability review and high-signal triage. - Spec and intent checks for generated pull requests. - Trust scoring for code, agent workflows, skills, tools, and MCP servers. - Policy gates that route changes to block, review, or ship decisions with audit evidence. ## What we publish here This Hugging Face organization is the public home for Cognium AI and security artifacts: - Reproducible benchmark summaries for semantic SAST and AI-assisted verification. - Demo Spaces that make scan, trust-score, and policy-gate outputs easier to inspect. - Dataset cards for public benchmark corpora and result metadata. - Research notes around AI-generated code security, agent trust, and model-assisted verification. ## Start here | Resource | Use it for | | --- | --- | | [cognium.dev](https://cognium.dev) | Open-source semantic SAST engine, installation, CI examples, and developer documentation. | | [Benchmarks](https://cognium.dev/benchmark) | Static-analysis benchmark results with reproduction notes and source links. | | [circle-ir](https://github.com/cogniumhq/circle-ir) | MIT-licensed static-analysis library and benchmark harnesses. | | [cognium.net](https://cognium.net) | Enterprise trust infrastructure for AI-generated code, agent governance, and policy gates. | ## Current focus We are building infrastructure for teams that use coding agents in real software delivery: - Security teams need verified, reproducible vulnerability evidence. - Engineering leaders need generated pull requests to arrive with behavior and risk summaries. - Platform teams need trust controls for the skills, tools, and MCP servers agents can use. - Regulated organizations need audit trails that survive procurement, compliance, and incident review. ## Contact For pilots, partnerships, or research collaboration, start at [cognium.net/pilot](https://cognium.net/pilot/) or open a discussion in the relevant public repository.