Please do not open public issues for security vulnerabilities.
If you find a security issue, email the maintainer privately. We will acknowledge within a reasonable timeframe, investigate, and coordinate disclosure.
Nursery is an agent-spawning framework. The highest-risk areas are:
- Secrets handling — leakage of OAuth tokens, API keys, credentials
- Container escape — an agent breaking out of its container onto the host
- Cross-agent leakage — one agent reading another agent's workspace or secrets
- Supply chain — compromise of the base image or dependencies
- Agents are assumed to be mutually untrusted. One agent should not be able to read another's state.
- The host gateway is trusted. Compromise of the gateway compromises all agents.
- Workspaces contain sensitive user data (memory, conversation history). Treat them like personal diaries.
- Issues in third-party model providers (Anthropic, OpenAI, Google, etc.)
- Issues in specific host runtimes (OpenClaw, Hermes) — report those upstream
We ask that you give us a reasonable window to address issues before public disclosure. Standard industry practice (90 days) is fine.
Thank you for helping keep this project safe.