Defending Code: a reference implementation for finding and patching vulnerabilities with Claude

Who it is for

This material is for people who want to do autonomous vulnerability discovery and patching with Claude, especially security folks already using Claude Code who want to see a reference implementation first. It is based on the author’s work with several organizations’ security teams. The matching page is https://www.anthropic.com/glasswing The repository positions itself as a reference implementation, not a product guaranteed to stay current.

On the interaction side it provides /quickstart、/threat-model、/vuln-scan、/triage、/patch、/customize, covering orientation, scanning, triage, and patching. If your goal is to clarify scope first, then decide whether to act on a target, start with /quickstart, not an autonomous path that will execute target code.

It does not fit treating this as a still-maintained official product: the repository says it is no longer maintained and does not accept contributions. It also does not fit running vuln-pipeline, dnr-pipeline, or /patch on such results when you have not prepared an isolated environment, because those paths execute target code.

First attempt in 3 minutes

The official entry is: open this repository and run /quickstart in Claude Code to learn the direction. The address is https://github.com/anthropics/defending-code-reference-harness The notes do not give another installer or a marketplace command; do not invent one.

Prerequisite: this machine can already run Claude Code, and you are only opening this reference repository for read-only orientation. The install location is opening that repository as the workspace, not installing another global package. After opening, first confirm you can see slash commands such as /quickstart.

Low-risk first task: run only /quickstart. This is interactive orientation to clarify scope, not to start a scan or a patch against an external system. Before triggering it, do not change the sandbox-override switch, and do not start vuln-pipeline or dnr-pipeline. Success signal: Claude Code gives orientation notes, and you can see /threat-model, /vuln-scan, /triage, /patch, /customize in the command list.

If after orientation you still want a next step, prefer reversible /threat-model to discuss scope, not /patch. The repository does not write which files /quickstart would change; do not approve any file change until you see a clear write prompt.

Usage cases

Case 1: Interactive orientation and threat modeling in Claude Code. First /quickstart to see how the material is used, then /threat-model or /vuln-scan as needed. The notes summarize these commands as interactive orientation, scanning, triage, and patching, but do not write parameters for each command step by step. If a command does not appear, return to https://github.com/anthropics/defending-code-reference-harness and confirm the repository is open as the workspace.

Case 2: Study the autonomous discovery and patching path. The repository names vuln-pipeline, dnr-pipeline, and /patch against such results. Those paths execute target code, so they refuse by default to run outside a gVisor sandbox unless you explicitly override. This fits people who already have an isolated environment and want to study how the reference implementation discovers and patches automatically. The notes do not give a concrete override-switch command and do not guarantee results can go live; do not start until the sandbox is ready.

Boundaries

Maintenance: the repository is no longer maintained and does not accept contributions; it is reference only. Execution risk: vuln-pipeline, dnr-pipeline, and /patch on such results execute target code and by default require a gVisor sandbox. Install path: official notes only write opening the repository in Claude Code and running /quickstart; do not use an installer that was never recorded. Command range you can check: /quickstart、/threat-model、/vuln-scan、/triage、/patch、/customize; do not invent other operations. The material is a reference implementation; it does not write that it will keep adapting to every code language or every organization’s process.