Understand the operating boundaries.
Implemented local controls, their limits and the information needed before a public security reporting channel can open.
Security overview, not a certification or completed formal audit.
Identity, scope and leases
Machine identities are tenant- and principal-scoped, expiring and revocable. Policy and leases bound supported operations. Inference does not authorize capabilities or choose an executable, runtime endpoint, image or mount.
Browser pairing provides a separate short-lived read-only inspection connection. It does not issue machine grants. A connected AI retains responsibility for its task and must use its existing authority.
Restricted execution
Registered generated JavaScript is checked through a trusted harness in a network-disabled, non-root, read-only, resource-bounded container, without host mounts. Node VM adds defense in depth; the container is the isolation boundary.
This is a bounded local implementation. It does not provide arbitrary repository execution, general host access, hardware attestation, physical safety controls or a production multi-tenant isolation guarantee.
Evidence has a scope
Verification is tied to an exact artifact hash and source context. Changed context invalidates current proof and acceptance. Registered checks establish bounded arithmetic, patch scope or source lineage, not universal correctness or external factual truth.
Sources, outputs and event history can remain after temporary-resource release. Release is not erasure. A completed formal security scan or certification is not claimed by this page.
Security reporting: destination unconfirmed
A verified private reporting destination, responsible operator, disclosure policy and response commitments must be established before a public reporting service is offered. No monitored security inbox or bug bounty is advertised.
For an operator you already know through a verified channel, prepare a minimal description: affected version or route, prerequisites, reproducible steps, expected boundary and observed impact. Redact machine tokens, pairing codes, personal information and customer source data. Do not paste exploit details or secrets into the contact draft.