
@lifeof_jer Those are execution-time conventions, not enforceable controls. If an agent can go from a normal question → find credentials → run a destructive API call, then the guardrails are observational, not preventive. Safety can’t rely on model behavior or in-context approvals.
English







