Skip to content

Security: rozmiarD/GovEngine

Security

SECURITY.md

Security Policy

GovEngine is an in-process governance kernel. It is not a scanner, exploit framework, sandbox, execution runtime, or remote authorization authority.

Supported versions

GovEngine source is currently the prepared/unpublished 1.0.0rc2 release candidate. Security fixes should target main until the stable release line exists. The public 1.0.0rc1 package has passed independent contract review. Final 1.0.0 promotion requires rc2 authentic review and publication, review covering that release commit and a completed candidate observation window without an open P0/P1.

Reporting issues

Please report security-sensitive issues privately to the project owner when possible. Do not publish credentials, private targets, raw runtime logs, or exploit details in public issues.

Security boundaries

GovEngine must not:

  • treat LLM prose as executable authority;
  • construct live shell commands from untrusted text;
  • violate the documented SCLite/GovEngine/RExecOp/profile dependency direction;
  • publish raw stdout/stderr, command logs, credentials, cookies, private paths, or private target identifiers;
  • claim authorization to test a target;
  • claim live vulnerability evidence from dry-run artifacts.

The detailed trusted computing base, attacker model and residual risks are in docs/THREAT_MODEL.md. Tested guarantees, digest bindings and explicit non-claims are in docs/SECURITY_GUARANTEES.md.

GovEngine runs in the host process. It does not claim resistance to a malicious or fully compromised in-process host that skips the kernel or ignores its decision.

Expected safe behavior

New safety-sensitive code should be deterministic by default, testable without live targets, and explicit about:

  • scope and policy decisions;
  • approved vs prepared execution shape;
  • dry-run/local/mock/live truth;
  • receipt/evidence non-claims;
  • bounded, recursively filtered metadata at key-resolution and trust-store adapter boundaries;
  • owner-review boundaries.

There aren't any published security advisories