Only the latest release receives security updates. Pin to a specific version for production and upgrade promptly when a security advisory is published.

Do not open a public issue.

1. Preferred: use GitHub Private Vulnerability Reporting.
2. Fallback: email security@latchgate.ai with subject [LatchGate Integrations Security].

Include: description of the vulnerability and its impact, affected package(s) and version(s), reproduction steps or proof of concept, and severity assessment if known.

We use CVSS v3.1 for severity assessment:

These are targets, not guarantees. Complex issues may take longer. We will communicate progress to the reporter throughout.

We follow coordinated disclosure. We ask reporters to keep findings confidential until a fix is released. We will coordinate a disclosure timeline with the reporter and request a CVE where applicable.

We credit reporters in the release notes and security advisory unless anonymity is requested.

We consider security research conducted in good faith to be authorized and will not pursue legal action against researchers who:

• Make a good-faith effort to avoid privacy violations, data destruction, and service disruption
• Only interact with accounts they own or with explicit permission of the account holder
• Do not exploit a vulnerability beyond the minimum necessary to demonstrate it
• Report the vulnerability through the channels described above before any public disclosure

This policy covers all integration packages in this repository:

• latchgate-integrations-common — shared discovery, schema conversion, result serialization, transport resolution
• latchgate-langchain — LangChain BaseTool integration
• latchgate-crewai — CrewAI BaseTool integration
• latchgate-openai-agents — OpenAI Agents SDK FunctionTool integration
• latchgate-pydantic-ai — Pydantic AI AbstractToolset integration
• latchgate-ai-sdk — Vercel AI SDK ToolSet integration
• CI/CD pipeline — GitHub Actions workflows, release pipeline

Security-relevant areas include output-only serialization (what the model sees), description redaction (enforcement topology exposure), identifier validation (path traversal), and the sync-to-async bridge.

• LatchGate core (server, kernel, auth, policy, ledger, providers) — report to latchgate-ai/latchgate
• Python SDK (latchgate on PyPI) and TypeScript SDK (latchgate on npm) — report to latchgate-ai/latchgate
• Example scripts in examples/
• Third-party framework bugs (LangChain, CrewAI, etc.)