daily-supply-chain-auditlisted

Run a comprehensive daily security audit of all installed MCP servers, Claude Code skills, agent configs, and AI CLI tooling. Detects supply chain attacks, tool-name shadowing, malicious skills, configuration drift, missing CVE patches, and untrusted package versions. Use this skill EVERY MORNING before starting work, on demand when installing a new MCP/skill, or after pulling Walter-OS updates. ALSO trigger when the user asks "is my agent setup safe", "audit my MCPs", "check for vulnerabilities", "any new CVEs", or mentions concerns about supply chain, tool poisoning, prompt injection, or malicious skills.
Xipher-Labs/walter-os · ★ 5 · AI & Automation · score 67

Install: claude install-skill Xipher-Labs/walter-os

# Daily Supply Chain Audit Audits the agentic toolchain end-to-end and refuses to let work continue if critical vulnerabilities are found. Designed for paranoia, calibrated for 2026's threat landscape (1,184 malicious skills found in ClawHavoc, CVE-2025-59536 hooks injection RCE, three vulns in Anthropic's own Git MCP, mcp-remote CVSS 9.6 RCE — supply chain attacks on AI tooling are real and ongoing). ## What it checks 1. **Tool versions** — `claude --version` ≥ 2.0.65, `codex --version` current, `gh`, `node`, `pnpm` not on known-vulnerable releases. 2. **Config drift** — `~/.claude/settings.json`, `~/.codex/config.toml`, and any `.mcp.json` in active repos diffed against signed baselines stored in `~/.config/walter-os/baselines/`. Unauthorized additions = block. 3. **Hooks integrity** — every hook in `~/.claude/settings.json` matches a sha256 in `~/.config/walter-os/hook-checksums.json`. New hooks require explicit operator approval (the script prompts). 4. **Installed MCP servers** — runs `mcp-scan` (Snyk) and `mcp-scanner` (Cisco) if available; queries `agentaudit.dev` and `mcpskills.io` for trust scores. 5. **Skills audit** — every skill in `~/.claude/skills/` and `~/.codex/skills/` gets static analysis: scripts shouldn't `curl | bash`, shouldn't write to `~/.ssh`, shouldn't egress to non-allowlisted domains. 6. **Tool definition drift** — stdio, HTTP, and SSE MCP servers loaded from `~/.claude/settings.json` are probed via `tools/list` and co