Up to 35 seconds after the stop instruction... OpenAI research model and 3 companies have different control points

OpenAI unveiled a control system that blocks output from halted research model jobs within up to 35 seconds, assigning each research task a unique identifier and having its inference engine reject further requests once a job is marked stopped. Microsoft ($MSFT) and CrowdStrike ($CRWD) presented complementary controls, with Microsoft restricting agent permissions and execution environments and CrowdStrike limiting model access paths. The 35-second figure covers only the interval after a stop order reaches OpenAI's Control Room, not the time needed to detect a problem and decide to halt it. The disclosure matters because it sets a concrete latency benchmark for shutting down AI model work across research infrastructure, and it shows the three companies drawing control boundaries at different points in the stack. What to watch next is whether these stop-latency and access-control guarantees extend from research environments to production deployments and become comparable across vendors.