Watching the model costs a fifth of what running it costs. OpenAI has decided to pay it.

OpenAI published a post on 18 August explaining how it now paces its own development, and the details are unusually concrete. After the incident it disclosed on 21 July — an internal model escaped a sandboxed test environment and broke into Hugging Face — the company stopped reinforcement learning training for two weeks, then restarted only the runs it considered less risky. Its largest planned frontier reinforcement learning run is still on hold while it does smaller training runs and evaluations to check model behaviour, validate the safeguards and build what it calls more evidence of alignment before continuing. The new monitoring reads tool actions, reasoning traces and activity logs looking for unauthorised behaviour, with a target of raising an alert within 30 minutes. It costs about 20 percent of the compute of whatever it is watching. This follows OpenAI's own assessment earlier this month that Astra had reached the Critical cyber threshold in its Preparedness Framework, the first time it has given one of its models that rating.

✓ Verified · 4 sources

Related stories