Security architecture

goodLAC's security model is an authority/effect boundary, not a claim that every deployment is vulnerability-free.

Key controls in the qualified path

• Default-governed Pi with explicit ungoverned bypass.
• Bubblewrap sandbox with constrained filesystem/process/network/environment authority.
• Controller-owned identity and capability validation.
• Deterministic standing policy and deny precedence.
• Exact one-request approvals.
• Policy re-evaluation immediately before dispatch.
• Durable emergency pause.
• Execution leases, idempotency, receipts, and durable effect truth.
• Credential separation from model context.
• Owner-only administration surface outside governed sandboxes.

Credential separation

Agents receive bounded effect interfaces, not service credentials. Credential-bearing integrations resolve credentials outside the model context. Request schemas and public receipt projections are designed to avoid carrying authority/credential material across consumer boundaries.

Fail-closed cases

Unknown capabilities, unsupported scope/material, malformed or stale authority objects, invalid identity bindings, unavailable required services, expired approvals/requests, inconsistent continuation state, and other invalid authority conditions close rather than silently broadening authority.

Security reporting

Do not publish undisclosed vulnerabilities in issues, discussions, or pull requests. Use GitHub Private Vulnerability Reporting through the repository's Security interface.

No response-time, remediation-time, bounty, disclosure deadline, or service-level commitment is promised unless ITECOB Inc. expressly agrees to one for a specific matter.