Defensive and Authorized Use Only
Linmas is designed exclusively for defensive security reviews of software, architectures, and systems that you own or have explicit authorization to inspect. Linmas must never be used for unauthorized offensive activity, exploit generation, or malicious testing against third-party systems.
Linmas is not a security certification body, does not provide legal compliance guarantees, and is not affiliated with any government entity or OpenAI.
Offline Replay vs Provider-Backed Review
Linmas enforces a strict distinction between offline deterministic execution and live provider-backed model calls:
- Offline Fixture Replay (Default Demo): Runs entirely on your local machine with zero network calls and no model invocation. Uses checked-in fixtures to validate policy logic, MCP tool behavior, and capsule construction.
- Provider-Backed Live Review: Opt-in review execution that calls model providers (e.g. Codex / OpenAI, Claude API). Requires explicit confirmation (e.g.
--live --yes) before any input bytes leave the local machine.
Data Transmission Boundaries
Linmas never transmits telemetry, background analytics, or codebase contents automatically:
- Only the explicitly scoped input artifact (e.g. the diff passed via
--input) is sent to the selected model provider during live review. - No surrounding uncommitted files, environment variables, or private API keys are gathered or transmitted.
- Provider authentication is managed directly by the host environment (e.g. Codex or user-managed API keys).
Human Review is Always Required
AI models can produce false positives, misunderstand domain-specific invariants, or miss subtle architectural flaws. Consequently, Linmas sets humanReviewRequired: true on all outputs. A passing policy status is an evidentiary checkpoint, not a replacement for engineer review.
Vulnerability Reporting
If you discover a security issue or vulnerability in Linmas, please report it responsibly according to ourSecurity Policy (.github/SECURITY.md) ↗.