Agent harnesses
defending-code-reference-harness
anthropics/defending-code-reference-harness
A reference implementation from Anthropic that pairs Claude Code skills with an autonomous, sandboxed vulnerability-discovery pipeline covering recon, finding, verification, reporting, and patching.
Topics
Related projects
cc-safety-net2kCC Safety Net is a pre-execution guard for AI coding agents that parses commands and blocks destructive Git/filesystem operations and secret access across many coding CLIs.
OpenAPPA1.6kA deterministic guardrail layer that sits between an agent and its tools, enforcing data-flow policies before every tool call so sensitive information cannot reach unauthorized destinations.
ClawVault1.1kClawVault is a security vault for OpenClaw and Claude Code that intercepts sensitive data, prompt injection, and dangerous commands through a proxy and agent hooks.
helm-d1kA DeepSeek Harness plugin that adds 49 security-analysis tools, 353 references, evidence-chain enforcement and switchable personas for authorized Android, web, native and malware work.
ClawKeeper1kHost-agnostic safety middleware for tool-using agents that blocks risky tool calls and delegates trajectory decisions to a Watcher.
maggy711Opinionated harness for Claude Code with skills, quality gates, memory, and multi-model routing dashboard.