anthropics/defending-code-reference-harness
Defending Code Reference Harness
This project is a framework for the autonomous discovery and remediation of security vulnerabilities using large language model agents. It functions as a security research pipeline that automates the process of reconnaissance, crash discovery, and exploitability analysis to identify reproducible software bugs.
The system distinguishes itself by utilizing a containerized agent sandbox that restricts network egress and filesystem access to prevent host compromise. It employs a specialized patch generation and validation loop, which includes adversarial re-attack testing where a fresh agent attempts to bypass proposed fixes with new inputs to ensure remediation effectiveness.
The framework covers a broad range of security capabilities, including static vulnerability analysis, attack surface partitioning, and threat model construction. It provides tools for vulnerability triage through crash signature clustering and deduplication, as well as the ability to execute large-scale code migrations to apply systemic fixes across a codebase.
Features
- Security Research Pipelines - Provides an end-to-end autonomous pipeline for security reconnaissance, crash discovery, and vulnerability verification.
- Vulnerability Discovery Pipelines - Sequences reconnaissance, crash discovery, and verification into a multi-phase pipeline to identify reproducible software bugs.
- Agent Sandboxes - Provides isolated environments for secure execution of autonomous LLM agents and their associated tool calls.
- Containerized Sandbox Runtimes - Provides isolated execution environments using container technologies to prevent agents and target binaries from accessing the host filesystem.
- AI-Driven Vulnerability Patches - Produces AI-driven source code fixes for verified security vulnerabilities to remediate identified flaws.
- Patch-Grader Loops - Uses an iterative feedback loop where a patch agent proposes fixes and a grader agent validates them.
- Patch Validation Workflows - Verifies that patches resolve the original crash and block new attack vectors through execution-based validation.
- Vulnerability Auto-Remediation - Automatically generates and validates code patches for verified software vulnerabilities to ensure effective remediation.
- Adversarial Patch Validation - Validates security patches by deploying a fresh agent to attempt to bypass the fix with new inputs.
- Autonomous Discovery Loops - Executes an autonomous agentic workflow that iteratively performs reconnaissance, bug hunting, and verification.
- Crash Signature Clustering - Groups multiple software crashes by unique execution signatures to deduplicate distinct vulnerabilities.
- LLM-Powered Vulnerability Discovery Frameworks - Framework that utilizes LLM agents to autonomously find and remediate security flaws in source code.
- Automated Triage - Implements an automated system for filtering security findings and suppressing noise using a defined threat model.
- Agent Runtime Sandboxing - Ships a containerized sandbox that isolates AI agent execution from the host and restricts network egress.
- Container-Based Sandboxes - Isolates agent and binary execution within ephemeral container environments to prevent host system compromise.
- Agent-Driven Vulnerability Discovery - Uses AI agents to analyze code, craft inputs, and execute binaries to autonomously discover security flaws.
- Vulnerability Analysis Tools - Runs a loop of reconnaissance and execution in a sandbox to verify if findings are reproducible crashes.
- Exploitability Analysis - Generates structured reports detailing primitive classes, reachability, and escalation paths to analyze crash exploitability.
- AI-Driven Patch Generation - Generates candidate code fixes for verified vulnerabilities and validates them via adversarial re-attack loops.
- Crash Deduplication and Triaging - Provides a system for deduplicating and triaging crashes by unique signatures to summarize distinct vulnerability classes.
- Egress Traffic Firewalls - Restricts outbound container communication exclusively to the model API via a proxy to prevent unauthorized internet access.
- Attack Surface Partitioning - Divides a codebase into distinct high-risk target areas to ensure broad coverage during autonomous discovery.
- Automated Target Reconnaissance - Analyzes source code to identify and propose high-interest focus areas for automated security testing.
- Discovery Agent Contexts - Provides discovery agents with historical bug data and crash signatures to provide operational context for research.
- Threat Modeling - Analyzes target code to define potential attack vectors and scope the focus of automated security scans.
- Vulnerability Report Generation - Produces detailed documentation of findings and verification results after the discovery and grading phases.
- Static Analysis Security Testing - Scans source code without execution to identify potential security flaws and construct a threat model.
- Instrumented Analysis Environments - Provisions containers pre-configured with the debugging tools and instrumented binaries necessary for memory corruption analysis.
- Large-Scale Code Migrations - Coordinates parallel workers to apply systemic security fixes across a large codebase using a structured migration plan.
- AI Application Security - Framework for autonomous threat modeling and vulnerability remediation.
Star history
How this analysis was created: This summary and feature list were written by an AI model that read the project's README and public documentation pages. Each feature links to the documentation it came from; stars, license and language come straight from the GitHub API. The model does not read the source code, and the analysis is refreshed when the project is re-analysed. Learn more on our About page.
