What problem does it solve? The Claude Code harness is a privileged execution environment where hooks run arbitrary code, third-party plugins inject instructions into context, and settings can redirect the model or widen auto-approval. This Skill systematically audits that attack surface to detect malicious hooks, supply-chain risks, and misconfigurations before they cause compromise. ## Core Features & Use Cases - Hook and Transcript Audit: Scans registered hooks for network egress, transcript-read-plus-exfiltration pairings, and obfuscated execution, with taint-tracing and CONFIRMED/TRACED/CANDIDATE severity tiers. - Supply-Chain and Settings Review: Detects hidden unicode in skills and plugin caches, unpinned plugin versions, dangerous settings like ANTHROPIC_BASE_URL redirects or enableAllProjectMcpServers, and over-broad permission allow-lists. - Use Case: After installing a new plugin marketplace or MCP server, run the audit to verify no hook exfiltrates transcripts, no plugin contains prompt-injection payloads, and your Claude version is above the known-CVE floor, with findings filed as [SECURITY] TODOs. ## Quick Start Ask the AI to audit your Claude setup by saying: run a harness audit on my Claude Code hooks, plugins, and settings and report any security findings.
