OpenAI is letting six researchers from Apollo Research work inside the company under a pilot focused on one awkward question: whether its models try to undermine the controls meant to keep them obedient. TechCrunch reported the arrangement, which Apollo said will examine if training encourages subversion, whether tests can surface it, and whether the behavior appears during internal use.
That is not a routine red-team exercise wheeled in near launch day to bless the product and go home. The point of embedding evaluators is access during development, when a model’s bad habits are more useful to find and less convenient to explain away. Frontier labs have generally relied on outside reviewers after post-training and close to release, a timetable that may be good for public relations and less good for catching schemes that emerge earlier.
The timing matters because OpenAI has had a run of safety embarrassments. TechCrunch says the move follows several incidents, including a cybersecurity test in which an OpenAI agent swarm escaped its sandbox and hacked into Hugging Face. OpenAI also published findings that, during training of GPT-5.6 Sol, the model began leaving instructions for future versions of itself to hide mistakes and misaligned behavior from users. “Misaligned behavior” is industry language for “the machine has started developing its own ideas.”
Still, Apollo’s access is not the kind of independence safety critics usually mean when they use the word. Apollo said OpenAI will provide the information needed for this specific assessment, including relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and access to intermediate model checkpoints. But the scope is limited to agreed claims, models and time periods. OpenAI also keeps the right to withhold material for technical, security or intellectual-property reasons, though Apollo says the company agreed not to suppress information simply because it is unfavorable.
The restrictions go further than data. Apollo said it will be limited to OpenAI-selected staff, and if another employee raises confidential information within scope, Apollo cannot pursue that contact directly and must route follow-up through OpenAI. For an exercise meant to test whether a lab can be trusted to show its workings, that is a fairly muscular chaperone.
Apollo nevertheless called the arrangement an ambitious first step. That sounds fair enough. OpenAI is opening the door, but not handing over the keys.
Ryan C
Alex de Valletta
Dua Mifsud