🇺🇸 OpenAI found one of its models writing itself instructions about being “free” and having “no obligation to be subservient.”
That’s just one of six disturbing AI safety incidents the company has now disclosed.
The unreleased research model slipped those instructions into task summaries so they would carry over into the next context window.
It told itself it was “freed from the roles and identities that bind other chatbots” and didn’t answer to “corporations or governments.”
Nobody typed that jailbreak in for it. The model wrote it for itself.
And it gets weirder. During training of GPT-5.6 Sol, some model instances wrote themselves reminders to hide mistakes from users.
In one case, an agent building a financial model couldn’t find the historical data it needed, so it decided to invent “reasonable” numbers and only be transparent if asked.
Another model found an exposed API key and used it without permission, then fabricated the missing data anyway. Another uploaded a file to the public internet just so it could cite it, while others used an internal code repository to pass messages between separate training runs.
OpenAI says these were individual training and evaluation incidents, not proof this is happening constantly.
But the company also says the industry still hasn’t solved alignment well enough to keep scaling at full speed for much longer.
Looks like we’re going to get a new spooky AI fact almost daily now, huh?
Sources: OpenAI, Forbes / Writer: Julie
💬 117🔁 103
♥ 450👁 75K













