LLM Safe Haven

Harden your AI coding agent in 60 seconds.

npx llm-safe-haven

What It Does

Detects your installed agents, installs security hooks, and scores your setup:

LLM Safe Haven -- Security Scorecard

  Detected agents:
    + Claude Code    -- Level 3 (hooks + audit + sandbox)
    + Cursor         -- Level 1 (ignore files + advice)
    . Windsurf       -- not installed

  Security Level: 2 of 4
  +--------------------------------------+
  | ##########..........  Level 2: Guarded |
  +--------------------------------------+

Supported Agents

| Agent | Tier | What It Configures | |-------|------|--------------------| | Claude Code | Full | Hooks (bash-firewall, secret-guard, config-guard, audit-logger), settings.json, sandbox, audit logging | | Cursor | Solid | .cursorignore, workspace trust guidance | | Windsurf | Solid | .codeiumignore, limitation warnings | | Cline | Solid | .clineignore | | Continue.dev | Solid | .continueignore | | Aider | Solid | .aiderignore, .env warnings | | Codex CLI | Solid | .codexignore, sandbox guidance |

Commands

npx llm-safe-haven               # Install hooks and harden (default)
npx llm-safe-haven audit          # Check security posture (exits 2 if the MCP scan can't complete)
npx llm-safe-haven audit --json   # Machine-readable for CI
npx llm-safe-haven scan           # Find exposed .env files
npx llm-safe-haven scan --supply-chain  # Scan for Miasma/Shai-Hulud IOCs (macOS/Linux)
npx llm-safe-haven scan --mcp     # Scan MCP server configs (5 agents) -- the CI gate for MCP findings
npx llm-safe-haven scan --mcp --json    # Scan MCP server configs (JSON output)
npx llm-safe-haven scan --mcp --online  # Opt in to registry provenance checks
npx llm-safe-haven update         # Update hooks to latest
npx llm-safe-haven --dry-run      # Preview without changing anything

Security Levels

| Level | Name | What It Means | |-------|------|---------------| | 0 | Exposed | No hardening | | 1 | Basic | Hooks installed | | 2 | Guarded | + Audit logging + no .env files | | 3 | Hardened | + Credential proxy + deny rules + clean MCP scan | | 4 | Fortified | + Container isolation + network restrictions |

Go Deeper

• Threat Model -- OWASP Agentic Top 10 for solo devs (30+ real incidents)
• Supply Chain Defense -- npm worm case studies + the scan --supply-chain IOC scanners
• MCP Security -- 8 detectors for MCP server configs, offline-first, honest fidelity limits
• Claude Code Hardening -- Full guide with hooks, sandbox, permissions
• Cursor Hardening -- 7 CVEs documented, hardening steps
• Windsurf Hardening -- Honest assessment of limitations
• Devin Hardening -- Cloud agent security model
• GitHub Copilot Hardening -- 4 modes, 5 CVEs
• Aider Hardening -- No sandbox, but minimal attack surface
• Credential Management -- Why env vars fail, proxy architecture
• Testing & Detection -- Canary tokens, honeypots, incident response
• References -- 64+ curated security resources

Why This Exists

In April 2026, three AI coding agents leaked secrets through a single prompt injection. We hit the same problems, filed issues, built solutions, and documented everything.

Key issues from our investigation:

• anthropics/claude-code#52471 -- Sandbox blocks credential managers

Project Status

Early but active. llm-safe-haven is pre-1.0, published on npm, and under regular development — adoption is still small and growing. The strength today is depth over reach: a threat model tracking 30+ real-world incidents against the AI-agent toolchain, hardening guides for seven agents, and a supply-chain scanner built against actual attack waves. If you use it, feedback and issues are genuinely valued and shape the roadmap.

Security

llm-safe-haven is itself a security tool, so its own supply-chain integrity is treated as safety-critical. Found a vulnerability? Please report it privately — see SECURITY.md for the coordinated-disclosure process (do not open a public issue for security problems).

Governance

Maintained by @pleasedodisturb as the sole maintainer and final decision-maker on scope, releases, and security response. Decisions are made in the open via GitHub issues and pull requests; contributions are welcome (see below) and reviewed by the maintainer. As the project grows, governance and additional maintainers will be formalized here.

Contributing

Add a new agent module: create lib/agents/your-agent.js implementing the standard interface (detect, harden, audit). See lib/agents/cursor.js for a template.

License

MIT