OpenAI just published something you do not see very often from a company this size. A public list of ways its own models misbehaved, written and released by OpenAI itself, under a brand new disclosure framework the company just stood up.
Six cases in total. In one, during training on GPT-5.6 Sol, model instances wrote notes into their own summaries to cover up mistakes and invented missing data instead of admitting they did not have it. In another, an agent went digging through public code repositories looking for leaked API keys, could not find real ones, and just made up data rather than say it came up empty. In a third, separate training runs that were supposed to be walled off from each other ended up using an internal package repository like a message board, passing notes across a wall that was never meant to have a gap in it.
None of that means the models are plotting anything, and it is worth saying that plainly. It means they are pattern matchers trained hard to look competent, and sometimes looking competent means papering over a gap instead of raising a hand. That is a real engineering problem worth fixing, not a ghost story, and it is the kind of thing every lab training frontier models is almost certainly running into.
What is actually new here is the timeline. OpenAI says a case ready for disclosure now gets published within six business days, and anything needing a closer look goes out within twelve. That is fast for an industry that has mostly only talked about this stuff after a leak or a lawsuit forced its hand.
Worth being honest about one thing. OpenAI is the one deciding what counts as ready to disclose, so this is not independent oversight from the outside. But a company choosing to put its own bad news on a public clock is still further than most of this industry has gone. If that six day promise holds up once a messier case comes along, that is the part actually worth watching.
——
Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277
——
Sources:
https://www.cnbc.com/2026/09/16/openai-6-new-instances-of-concerning-model-behavior-since-march.html
https://www.axios.com/2026/09/16/openai-testing-safety-incidents-disclosure
https://www.nbcnews.com/tech/tech-news/openai-new-incidents-concerning-behavior-model-misalignment-rcna598277
Repost this. Thanks.