Anthropic Releases Open-Source Framework for Autonomous Vulnerability Discovery
Summary
- • Anthropic open-sourced a reference implementation for autonomous AI-driven vulnerability discovery and remediation using Claude.
- • The framework covers a full recon-to-patch pipeline and is tied to the commercial Claude Security hosted product.
- • Developed alongside Anthropic's Claude Mythos Preview security initiative; compatible with Claude APIs on Bedrock, Vertex, and Azure.
Details
Open-source autonomous vulnerability pipeline released as reference implementation
Anthropic published defending-code-reference-harness on GitHub, providing a full autonomous security pipeline built on Claude. The harness is derived from real-world learnings gathered by partnering with enterprise security teams since the Claude Mythos Preview launch. It is intentionally a non-maintained reference — teams are expected to fork and customize rather than rely on upstream updates.
Pipeline covers five stages: recon, find, verify, report, and patch — pre-configured for C/C++ with ASAN
The reference harness is pre-configured to find memory safety vulnerabilities in C and C++ codebases using Docker containers and AddressSanitizer. The five-stage loop mirrors standard security research workflows, enabling teams to go from zero coverage to autonomous continuous scanning within roughly two weeks following the structured onboarding plan.
Autonomous code execution sandboxed by default via gVisor; override requires explicit opt-in
Static analysis commands only read and write files and carry no execution risk. The autonomous pipeline, which runs target binaries, defaults to requiring a gVisor sandbox and will refuse to proceed without one unless the operator explicitly overrides the setting — a deliberate safety guardrail for running untrusted code during automated scanning.
Claude Security is the companion hosted product for teams wanting a managed multi-project solution
For organizations that prefer not to operate their own pipeline, Anthropic offers Claude Security — a hosted service that scans repositories, applies a multi-stage verification pipeline to reduce false positives, and provides lifecycle management for findings including triage, fix validation, and rapid patch generation across multiple projects simultaneously.
Framework is API-agnostic: supports AWS Bedrock, Google Vertex AI, and Azure deployments
The reference harness is designed to work with any Claude API access path — including AWS Bedrock, Google Vertex AI, and Azure-hosted Claude endpoints — making it accessible to enterprise teams with existing cloud commitments and data-residency requirements without locking adopters into direct Anthropic API access.
Six Claude Code slash-commands enable interactive security workflows from within the IDE
The harness ships with six Claude Code skills: /quickstart, /threat-model, /vuln-scan, /triage, /patch, and /customize. Static skills are read-and-write-only and safe to run unsandboxed. The /patch and /customize skills involve executing code and editing harness configuration respectively and require the gVisor sandbox.
Product Launch = new tool or service released; Tech Info = implementation and architectural detail; Security Alert = safety or threat-relevant information; Infrastructure = deployment and integration context
What This Means
Anthropic is codifying enterprise security workflows into a reproducible, open-source artifact — lowering the barrier for security teams to adopt AI-assisted vulnerability research without starting from scratch. The dual-track strategy (open-source reference plus Claude Security managed product) mirrors the playbook used by developer-tooling companies to build ecosystem trust while monetizing at scale. For AI practitioners in security, the gVisor sandboxing requirement and the staged onboarding plan signal that Anthropic is thinking seriously about safe deployment of autonomous code-execution agents, not just capability demonstrations.