OpenAI slows down its own training

OpenAI listed what it changed. Workload isolation: stronger sandboxes, closed-off environments, for anything that runs code written by models or other untrusted code. The same applies to software that could be compromised while it processes model output. Network isolation: controls designed so that breaking into one workload does not by itself give access to the internet or to internal networks. Continuous security testing: vulnerable shared services were removed, permanent access rights were cut back, and OpenAI now uses its own models to attack these boundaries all the time.

OpenAI also expanded the monitoring of the chain of thought, the reasoning steps a model writes down, across its advanced models. It strengthened alignment training and evaluations. It is updating its Preparedness Framework, its internal safety policy, to bring monitoring, alignment and containment together.

On 7 August OpenAI had already put monitoring for risky actions and misalignment on every agentic use of Astra, meaning every use in which Astra acts on its own. Before the release it also said publicly that it could not rule out that Astra was cyber critical.

In September, after a series of disclosures, OpenAI paused training on its most capable models. It said it would resume "only when we are confident that we have additional safeguards and alignment improvements in place".

Some people say that nothing is being done. Against that view, I find this the best argument in the dossier: a company that slows itself down in public and lists what it changed.

But OpenAI imposed the pause itself and chose its own timing. Nothing obliged OpenAI to pause, and nothing stops it from starting again. Compare the American bills, which would make a pause a legal obligation, and the European direction, which went the other way in May.