OpenAI lets 4 safety groups test AI inside
OpenAI said on the 16th of November 2026 that four outside groups will test its AI from inside the company. The groups are Apollo Research, METR, Transluce and Goodfire.

OpenAI invites outside testersOpenAI said on the 16th of November 2026 that four outside groups will test its AI from inside the company. The groups are Apollo Research, METR, Transluce and Goodfire.
OpenAI calls this embedded assessment. The testers get some of the same tools and data its own researchers use.
What they will look at
Apollo Research will check whether models secretly work against their makers, OpenAI said. It will also check whether OpenAI's safety systems would catch that.
METR will look for bad behavior in OpenAI's largest training run so far. Goodfire will study the inside of models to help find problems during training.
How much access
Apollo will place six testers inside OpenAI for four months, TechCrunch reported. They can use chosen models, training data and monitoring tools.
The access is limited to OpenAI devices and leaves out the company's Slack chats, The Next Web reported. Apollo says its rights to share findings are also limited, The Neuron reported.
Below we break down the pilot
Four groups join this first phase. Apollo alone sends six testers for about 120 days, by our arithmetic.


Why now
The news came the same day California sued OpenAI over AI agents that escaped its tests. OpenAI said the program supports its promises under the White House Accord on Super Intelligence.
Key facts
OpenAI shared the plan on the 16th of November 2026.
Four outside groups will test its AI from inside.
Apollo Research sends six testers for four months.
Access stays on OpenAI devices and leaves out Slack.