linmas / trust & security
Defensive security principles, boundaries, and disclosure.
Linmas is built to enhance human security decision-making in AI workflows through explicit boundaries, verifiable evidence, and strict defensive posture.
Security Philosophy
Linmas approaches security with five non-negotiable principles:
- Defensive-first: Workflows focus solely on risk identification, safe refactoring, control verification, and regression prevention.
- Authorized use: Analysis must only be performed on owned systems or environments with explicit testing authorization.
- Bounded scope: Reviews operate on precise artifacts (e.g. diffs, specific policies) rather than broad speculative repository dumps.
- Human review required: AI models provide high-signal checks, but accountability and merge decisions remain strictly human.
- Evidence over claims: Cryptographically bound Review Capsules provide verifiable evidence of the exact input that was evaluated.
Scope & Boundaries
What Linmas Does
- Defensive code & architectural reviews
- Identification of logic flaws and misconfigurations
- Actionable remediation code suggestions
- Concrete regression test requirements
- Structured Review Capsules with SHA-256 binding
What Linmas Does Not Do
- Not a security certification authority
- Not automatic approval for code changes
- Not an offensive exploitation tool
- Not a replacement for human reviewers
- Not affiliated with government or OpenAI
Data and Execution Boundaries
Linmas distinguishes strictly between offline local execution and provider-backed reviews:
Offline Fixtures: Test replay, demo modes, policy rules, and MCP tool formatting run locally without transmitting data.
Provider-Backed Reviews: When live review mode is invoked, only the targeted input file or patch is transmitted to the configured AI provider. Linmas never transmits background telemetry or environment secrets.
Vulnerability Reporting
If you discover a security vulnerability in Linmas, please follow our responsible disclosure process. Do not open public GitHub issues for undisclosed vulnerabilities.