Overview

docs / 01_overview

← all docs

Overview

Part of the DCRH doc set — see 00_START_HERE.md.

What it is

DCRH (Defending Code Reference Harness) is Anthropic's open-source reference implementation for autonomous vulnerability discovery and remediation with Claude — packaged learnings from partnering with security teams at several organizations (the "Glasswing" work). It is a crash-testing crew for source code: recon agents map the attack surface, isolated find agents craft malformed input until a memory-safety-instrumented build crashes reproducibly, a fresh-container grade agent re-verifies the crash without seeing the finder's reasoning, a judge checks for duplicates, a report agent writes the exploitability analysis, and a patch agent drafts a fix that a grading loop tries hard to disprove before trusting it. A parallel detection & response track applies the same verification-first discipline to log corpora instead of source.

Key concepts

Two halves

  1. Interactive skills (.claude/skills/): /quickstart, /threat-model, /vuln-scan, /triage, /patch, /dnr-hunt, /dnr-respond, /customize (+ contributor-only /verify).
  2. Autonomous pipelines: harness/ (vuln-pipeline, ~4,381 LOC) and dnr_harness/ (dnr-pipeline, ~876 LOC), both driven through bin/vp-sandboxed.

Where to go next

Related: 00_START_HERE.md.