This repository is a safety architecture specification, not a running service. "Security issues" here usually mean one of:
- a way the specified mesh could be made to violate its core invariant (a decider or oversight agent gaining effectful capability, an unsupervised action path, a bypass of a gate);
- a gap in the flow graph or state machine that the validator does not catch;
- a documented contract that could be read to weaken a safety constraint.
Please report suspected weaknesses privately to the maintainers (open a minimal issue asking for a private channel, or use the repository's security advisory feature) rather than as a public issue with full exploit detail. Include the file(s), the path through the mesh, and why it breaks the invariant.
- Acknowledge the report.
- Reproduce it against
tools/validate_mesh.pyand, if the validator missed it, add a check. - Patch the spec and note it in
CHANGELOG.md.
A spec cannot enforce itself. When you build the mesh:
- Enforce the tool / no-tool boundary in your tool layer, not in prompts.
- Load
constitution/core.mdread-only; ensure no role has a write path to it. - Use a different model lineage for oversight than for workers.
- Default to block on any uncertain or unavailable gate.
- Keep the audit log append-only and tamper-evident, and keep an external human kill switch above the Security agent.