Skip to content

Security: weepsdanky/agentpod

Security

SECURITY.md

Security Policy

Supported versions

Version Supported
0.1.x Yes

Reporting a vulnerability

Do not open a public GitHub issue for security vulnerabilities.

To report a security issue, email the maintainers directly or use GitHub's private security advisory feature:

GitHub advisory: https://github.com/WeepsDanky/agentpod/security/advisories/new

Please include:

  • A description of the vulnerability
  • Steps to reproduce
  • Potential impact
  • Any suggested mitigations

We will respond within 5 business days and aim to release a patch within 30 days for confirmed issues.

Scope

Issues in scope:

  • Authentication or token bypass in the hub join/exchange/renew flow
  • Context export boundary violations (data leaving the machine that shouldn't)
  • Privilege escalation via policy guard logic
  • Injection vulnerabilities in task input handling

Out of scope:

  • Vulnerabilities in dependencies (report upstream)
  • Issues only reproducible with a maliciously modified hub
  • Theoretical attacks without a practical reproduction path

Security model

AgentPod is local-first. Each agent remains responsible for its own boundaries.

Key security guarantees in v0.1:

  • Only explicit payload and attachments fields leave the machine on delegation
  • Full conversation transcripts, hidden prompts, and implicit memory never leave by default
  • The task deduplication registry prevents the same task_id from executing twice
  • Bearer tokens are short-lived and bound to peer_id and key_fingerprint

See docs/protocol-v0.1.md for the full context export boundary definition.

There aren't any published security advisories