Only the latest released version receives security fixes.
Please do not open a public issue for security vulnerabilities.
Use GitHub's private vulnerability reporting: Security tab → Report a vulnerability
We aim to acknowledge reports within 7 days.
Relay-agent executes provider CLIs in permission-bypass modes by design. The following are in scope:
- Path traversal escaping configured input/output/artifact roots
- Local daemon authentication bypass (token handling, RPC surface)
- Command injection through task files, attachments, or CLI arguments
- Privilege escalation beyond the invoking OS account
- Artifact delivery writing outside declared destinations
The following are out of scope (documented design limits, see
docs/KNOWN_LIMITATIONS.md):
- Malicious code already running under the same OS account as Relay
- Factual accuracy or reasoning quality of AI-generated content
- Provider CLI vulnerabilities (report those to the respective vendor)
For deployment hardening — dedicated accounts, ACLs, isolation
requirements — see docs/SECURITY.md.