Put the control where every request has to pass: Control Room, Nitro Sandboxes and log scanning
Three announcements this week solve different problems with the same move: they put a control where it cannot be skipped by the code it is meant to govern.
OpenAI's Control Room, described on its site, is the clearest case. Research workloads change constantly and are written by many teams, so OpenAI did not ask them to implement a stop. The check lives in the few inference code paths every workload must use: each request carries a job ID, the engine asks a central service for permission, and a stopped job gets no more tokens. The service fails closed, so switching it off stops all research inference rather than all checks (OpenAI).
AWS Nitro Sandboxes, announced on 9 November, do the same for isolation. Instead of trusting a container runtime and the code inside it, each EKS pod or ECS task runs in its own Nitro virtual machine, and secrets can be released only after cryptographic attestation of what is running (AWS). The boundary is in the hypervisor, which the workload cannot reconfigure.
GitHub's new secret scanning for Actions workflow logs moves a check to the output. Masking only hides values Actions knows to be secrets; a scan of the finished log catches what the workflow author never declared (GitHub).

The limits each one states
None of these is complete, and each source says where it stops. Control Room has a 35-second worst case because authorisation decisions are cached. Nitro Sandboxes cover most Nitro instance types, not all. GitHub's preview does not scan old logs, ignores custom and generic patterns, and sends no notifications, so an alert nobody looks at protects nothing.
What to take from it
The pattern is old: a circuit breaker does not rely on every appliance behaving. What changes is where teams draw the line. If a safety property depends on each workload, pipeline or pod doing the right thing, it will eventually fail on the one that does not. Finding the narrow shared path, the engine, the hypervisor, the log store, and enforcing there, is cheaper than auditing every caller.