Separate the responsibilities
Collection, investigation, policy, approval and execution are separate responsibilities in the proposed architecture. The investigating language model should not hold unrestricted customer credentials or approve its own actions.
Every connection needs an identity
The design calls for per-tenant separation, distinct collector identities, short-lived or revocable credentials where supported and narrowly scoped connector permissions. Adding tenant_id to a table is not, by itself, proof of isolation.
Minimise what is collected
Security telemetry may contain sensitive information. Collection scope, redaction, retention, access and model-processing rules need to be explicitly documented and configurable. The website demonstrations use no real telemetry.
Make execution verifiable
A requested action needs an exact resource, authorised actor, policy version, time limit, idempotency controls and an observable outcome. Commands sent without an acknowledgement cannot be reported as successful containment.
Treat updates as a trust boundary
Signed releases, protected build pipelines, dependency review and rollback are design requirements for any future privileged collector. No downloadable endpoint agent is supplied on this website.
Validate before production claims
Operational release requires threat modelling, isolation tests, collector review, load testing, failure handling, false-positive measurement and independent assessment. No certification or completed assessment is claimed here.