Anthropic has open-sourced a cybersecurity reference framework capable of achieving over 50% success on finding critical code vulnerabilities, compared to under 5% for traditional baseline detectors. Based on security learnings from the company's Claude Mythos Preview, the repository automates an end-to-end security workflow spanning discovery, verification, and patch generation. Developers and security teams can now leverage this pipeline to systematically audit source code for severe software flaws.
Autonomous Discovery Across Seven Pipeline Stages
The repository, titled defending-code-reference-harness, establishes an autonomous loop designed to discover and remediate software flaws with Claude. The process operates through seven structured execution stages:
- Build & Recon: Compiles target software with AddressSanitizer (ASAN) and uses a lightweight agent to partition the codebase into isolated attack surfaces.
- Find & Verify: Parallel subagents generate malformed inputs to trigger crashes, which a separate grader agent must reproduce in a fresh container.
- Dedupe & Report: A judge agent filters out duplicate reports, allowing a report agent to draft exploitability analyses covering reachability and severity.
- Patch Generation: A patch agent proposes code fixes, which are validated against existing test suites and subjected to re-testing by fresh find agents.
To protect host environments during execution, the framework enforces isolation through gVisor sandboxing. The autonomous pipeline executes target code inside containers with egress strictly restricted to Anthropic's API.
Interactive Skills and Multi-Stage Triage
In addition to fully autonomous runs, the repository includes interactive Claude Code skills designed for developer-guided workflows. Commands like /threat-model, /vuln-scan, and /triage allow security engineers to manually inspect repositories and rank static findings before generating candidate fixes.
During benchmark testing on complex vulnerability targets, the framework achieved over 50% success on its hardest test cases, compared to under 5% for a baseline detector. For organizations seeking a managed service rather than self-hosted infrastructure, Anthropic also offers Claude Security to scan repositories and validate fixes across multi-project environments.
Limitations and Operational Requirements
While the harness provides robust defensive patterns, Anthropic explicitly notes that the repository is an unmaintained reference implementation and does not accept external contributions. Furthermore, the pipeline is pre-configured out of the box specifically for C/C++ memory vulnerabilities using Docker and ASAN, meaning teams working with other programming languages must manually port the logic using the skill.