Start Here — Defending Code Reference Harness (DCRH) Onboarding Guide
Welcome. This documentation set explains DCRH
(anthropics/defending-code-reference-harness) — Anthropic's open-source reference
implementation for autonomous vulnerability discovery and remediation with Claude. It is
not maintained and not accepting contributions; it exists to be read, forked, and
ported, not depended on as a service. The managed equivalent is Claude Security.
Pick your path
| You are… | Read, in order |
|---|---|
| Non-technical (PM, analyst, manager) | architecture.html (Simple mode) → Overview → the "Explain-it-simply" boxes throughout |
| Security analyst | Overview → Data Flow → Sandboxing & Trust Boundaries |
| Junior developer | HLD → Data Flow → Module Flow → LLD |
| Reviewer / architect | HLD → LLD → Graph Intelligence Findings |
| Operator / hands-on user (want to run it) | Process & Runbooks |
What's in this folder
Anthropic -Defending Code Harness/DCRH/
├── docs/
│ ├── 00_START_HERE.md ← this file
│ ├── 01_overview.md ← what DCRH is, key concepts
│ ├── architecture.html ← interactive explainer (Simple/Technical toggle, module flow, pipeline stepper)
│ ├── data-flow.html ← rendered data-flow diagrams (Mermaid)
│ ├── architecture-overview.md ← single-file technical reference (superset, for printing/sharing)
│ ├── design/
│ │ ├── HLD.md / HLD.html ← High-Level Design
│ │ └── LLD.md / LLD.html ← Low-Level Design (incl. deterministic-vs-LLM breakdown)
│ └── modules/
│ └── README.md ← module index (points into architecture.html's Module Flow)
├── diagrams/ ← Mermaid sources (.mmd) + module-graph.html
├── tool-outputs/ ← real Graphify + CodeGraph exports, module_graph.json
└── tools/ ← the generators (never hand-edit the .html outputs; edit these, regenerate)
Long-term notes (Overview, Pipeline Stages, Interactive Skills, Sandboxing, Detection &
Response, Glossary, Process & Runbooks, Project Log) live in the paired Obsidian vault:
Defending Code Harness Vault/DCRH/, starting at DCRH — Map of Content.md.
The one-paragraph summary
DCRH runs a detect → verify → report → patch loop for memory-safety bugs in C/C++: build the target with ASAN, spin up parallel find agents to craft a 3-of-3 reproducing crash, re-verify it in a clean container that never saw the finder's reasoning, judge it against known bugs, write and score an exploitability report, then draft and adversarially grade a fix. A parallel detection & response track runs the same verification-over-vibes discipline against logs instead of source: hunt for an attacker already present, scope the blast radius, propose a response. Every agent-spawning subcommand runs inside a gVisor container with network egress restricted to the model API — enforced in code, not in prompts.
⚠️ The autonomous pipelines (
vuln-pipeline,dnr-pipeline) execute target code and refuse to start outside the gVisor sandbox unless--dangerously-no-sandboxis passed. The interactive skills (/quickstart,/threat-model,/vuln-scan,/triage) only read/write files and need no sandbox;/patchand/customizecan edit code — review before approving.