Security model

Authority is constrained by design.

GATOCP separates public presentation, AI reasoning, identity, policy decisions, commercial entitlements, and privileged execution so no single layer silently becomes trusted.

Boundary

Private control plane

The public website exposes no operational API or private runtime data. Sensitive control paths remain separate.

Identity

Scoped, short-lived, revocable

The roadmap moves toward workload identity, trust domains, rotation, revocation, and mTLS-ready credentials instead of shared permanent secrets.

Policy

Decision is separate from enforcement

Policy decisions are designed to be attributable and fail-closed; executors enforce decisions rather than self-authorizing.

AI

No unrestricted root or shell

AI produces bounded interpretation and plans. Availability of a model or agent never implies execution authority.

Workflow

Durability without duplicate side effects

Long-running operations evolve toward checkpoint, idempotency, retry, timeout, compensation, and deterministic recovery.

Supply chain

Provenance and attestations

Source, build, artifact, deployment, and verification evidence evolve toward independently verifiable cryptographic provenance.

Permanent invariants

Higher intelligence does not get higher authority automatically.

Tenant isolation, capability truth, approvals, kill switches, verification, rollback, and human override remain independent of model quality or roadmap maturity.

Tenant isolation

Future multi-tenant scale requires negative isolation tests across data, credentials, actions, telemetry, memory, jobs, and evidence.

Recovery before autonomy

Backup, restore, rollback, independent off-host recovery, and measured RPO/RTO must be proven before broader execution authority.

Capability flags are not permissions

Feature activation and commercial entitlement cannot override identity, policy, approvals, or runtime capability truth.

External protocols stay behind governance

Connectors, MCP, A2A, and future integrations remain adapters; none may bypass native GATOCP controls.

Security contact

Report security concerns through the published channel.

The canonical disclosure entry point is /.well-known/security.txt. Do not include credentials, private keys, or production secrets in an initial report.