OpenAI moved its model emergency brake into the inference layer. Shutdown takes at most 35 seconds

OpenAI has deployed Control Room, which checks research workload requests and halts further model output within 35 seconds after an intervention decision. Enforcement sits in the inference layer rather than mutable experiment code.

Every research workload receives its own pass to the model

Every agentic inference workload needs an approved safety case and job ID. The inference engine asks Control Room whether that ID may continue. Once stopped, code can keep running but receives no further model output. OpenAI says the system fails closed, denying inference when Control Room is unavailable.

Safety enforcement moves from experiments into shared infrastructure

Research code changes quickly and every team assembles it differently. An inference-layer check creates one shared point that each workload crosses for another answer. For agent operators, that pattern matters more than the product name because it separates monitoring from enforcement and is easier to audit.

The button works only after a monitor makes a decision

The stated 35 seconds run from an intervention decision to halted output. They say nothing about detection speed or decision accuracy. The account also comes from OpenAI, so independent evidence on availability, false alarms and bypass attempts is missing. Control Room removes inference access rather than killing the whole process.

Incidents will test whether the central brake survives its own failure

Real incidents must show coverage, network-failure behavior and whether a workload can evade the required job ID. Published red-team results for this layer would matter too. The architecture is sensible, but safety infrastructure earns trust when something breaks.

Lilith's verdict

An operator can pull the brake, but a monitor must identify the right train and every wheel must pass the checkpoint. Control Room earns its keep when the track supervising everything else fails.