comis-dev-crew is pre-release software under active development. No version is supported for production use, and no security guarantees are offered for any build. Do not run it against repositories, credentials, or hosts you are not prepared to lose access to.

None. The repository has no tagged release. Security fixes land on main only.

Report privately. Do not open a public issue for a suspected vulnerability.

Use GitHub private vulnerability reporting: https://github.com/comisai/comis-dev-crew/security/advisories/new

Please include the affected commit, the component (devcrew-service, devcrew, devcrew-mcp, devcrew-report, or a shared adapter), a reproduction, and the impact you believe it has. Redact credentials, tokens, socket paths, and worker output from anything you attach.

Expect an acknowledgement within 7 days. Because this is a pre-release project maintained without a funded security team, remediation timelines are best effort and will be agreed with you on the advisory thread.

The following are in scope and treated as security-relevant:

• Authority escalation across the local API, MCP facade, or reporter seam, including any path where worker or model content grants capability.
• Escapes from filesystem containment: worktree, runtime root, attachment mount, or approved-root resolution, including symlink and TOCTOU variants.
• Credential exposure in tracked files, arguments, logs, reports, events, or diagnostics.
• Introduction of a second durable writer to the SQLite store, or a replay or reconciliation defect that turns an unknown outcome into a claimed success.
• Command execution reachable from worker, model, forge, or protocol content.

The following are out of scope:

• Findings that require an already-compromised owner account, since the local API, database, and runtime directories are owner-only by design.
• Missing hardening in code paths explicitly marked as not yet implemented, including the protected live campaign and unratified deferred stages.
• Reports produced solely by automated scanners without a demonstrated impact.

The threat model and the invariants this project refuses to weaken are recorded in AGENTS.md, primarily its Security section. A report showing that a stated invariant does not hold in the implementation is always welcome, even when the practical impact is limited.