Anthropic has released the "defending-code-reference-harness," an open-source reference implementation designed for autonomous vulnerability discovery and remediation using Claude. The harness facilitates a complete security lifecycle: reconnaissance, finding, verification, reporting, and patching. Currently configured for C/C++ memory vulnerabilities, the system utilizes Docker and AddressSanitizer (ASAN) to detect crashes. The autonomous pipeline operates through seven distinct stages: building the target into a Docker image, performing reconnaissance to partition subsystems, employing parallel agents to find crashes, using a grader agent to verify proof-of-concept (PoC) execution, deduplicating findings via a judge agent, generating structured exploitability reports, and validating patches through automated build and regression testing. To ensure safety, autonomous pipelines execute within gVisor sandboxes with restricted egress. Beyond proactive scanning, the repository includes "detection and response" (D&R) capabilities, such as the /dnr-hunt and /dnr-respond skills, intended to identify attackers within system logs and propose incident response plans. While the harness is a reference implementation rather than a finished product, it provides a template for customizing security pipelines for various programming languages and vulnerability classes by redefining signal types and build processes.