OpenAI says its Control Room system can block any further output from a research model within 35 seconds of receiving a stop instruction. Authorization is checked by the inference engine through a job identifier rather than by the code of each individual research workload. The delay begins only after the instruction is received, as problem detection, investigation and the decision to intervene occur beforehand. Once the job is stopped, the engine denies its requests, while the process may remain active and continue serving other authorized jobs. OpenAI says the mechanism is already deployed across its research infrastructure and that remaining workloads are progressively adopting a fail-closed mode that denies inference requests if Control Room becomes unavailable.
Source: Read the original article
