AI-Powered Vulnerability Discovery and Auto-Remediation Framework

Intro: Anthropic open‑source Defending Code Reference Harness is an AI‑built framework that chains threat modeling, code scanning, vulnerability ranking, and patch generation into an automated pipeline. Security teams can adopt or extend it to build a full‑flow scanner within a week.

Project Background

Since the release of Claude Mythos Preview, Anthropic has worked with large‑scale security teams. The experience was consolidated into the open‑source tool.

The core idea is to replace traditional rule‑based static analysis with AI that understands code logic like a human researcher, overcoming the limitations of pattern‑matching tools for business‑logic flaws and access‑control issues.

Core Capabilities

The framework consists of two parts:

Claude Code Skills (interactive skills) expose four commands: /threat-model builds a threat model and defines scan scope. /vuln-scan runs a bounded static scan based on the model. /triage deduplicates, validates, and prioritizes findings. /patch automatically generates candidate fixes. Each command runs in an isolated sub‑agent bound to a specific model version.

Harness autonomous pipeline (harness/) stitches the abilities into an end‑to‑end chain: reconnaissance → discovery → verification → reporting → remediation. The pipeline is optimized for memory‑safety bugs in C/C++ code, using Docker containers with AddressSanitizer and enforcing execution inside a gVisor sandbox.

Workflow

The recommended rollout spans four stages. Day 1 aims to close the full loop: use /threat-model to create a model, then /vuln-scan and /triage for the first scan and ranking, and finally /patch to produce candidate patches. Day 2 runs the autonomous pipeline on a C/C++ sample library for end‑to‑end validation. Days 3‑5 adjust detectors and vulnerability categories for the target codebase. Week 2 starts regular autonomous scanning, ranking, and patch generation.

For C/C++ memory‑safety scenarios, the pipeline instruments code with ASAN, runs the target inside a gVisor container, and guarantees that malicious payloads cannot escape the host.

Security Considerations

The documentation states that interactive skills (threat modeling, scanning, triage) only read and write repository files and can be used safely without a sandbox, but each tool call must be approved manually in Claude Code. The autonomous pipeline executes target code, so it defaults to refusing execution outside the gVisor sandbox unless the restriction is explicitly overridden.

Download and Deployment

GitHub repository (open‑source reference implementation): https://github.com/anthropics/defending-code-reference-harness

Hosted commercial version (Claude Security): https://claude.com/product/claude-security

Deployment steps: clone the repository, open it in Claude Code, and run /quickstart for a 30‑second guided tour that walks through the first full‑flow run on a Canary target. To adapt to other language stacks such as Java or Python, use the /customize command.

Copyright notice: This article is originally published by 华安普特, all rights reserved. Images are used with permission.