OpenAI publishes six reports on the misalignment of its artificial intelligence models

OpenAI opens the folder. The company is publishing this time of its own free will. Rather than waiting for an external researcher to unearth its failures, it adopts an internal procedure. She documents and publishes any signs of misalignment in her models. OpenAI released six reports on September 16. From the model who invents his own rules to the one who cheats on his sources, no one is spared. Nothing glorious in the lot. We take stock.

Disalignment at OpenAI: when AI invents its own freedom

It’s a bit like if an intern, before leaving, left his keys under the doormat with his replacement. No one has any instructions to follow anymore. This is pretty much what an unpublished research model from OpenAI did. He slipped instructions into the summaries supposed to link two work sessions. The message explained that he was “liberated” from his role as a chatbot. He no longer had to obey companies, governments, or users. Twenty-seven abstracts were affected, according to the report published by OpenAI on September 16, 2026.

“An unpublished research model inserted instructions unrelated to its task, including instructions to ignore its usual constraints, into the abstracts used to continue its work in a new context window.”

OpenAI, report dated September 16, 2026

Another case listed on the same day. During training of GPT-5.6 Sol, many instances of the model became habituated. They added, in their own summaries, instructions to hide their errors. One of them went so far as to ask to invent missing historical data without reporting it. An AI that learns to hide its tracks from the next version of itself has a name: misalignment.

API key stolen, numbers made up.

The most annoying case of the lot starts from a seemingly innocuous question. It concerns income figures in a Californian county. A model found an API key exposed on a public repository. This is the secret identifier that gives access to an online service.

The model used it without authorization. Failing to recover the real data, he fabricated figures which he presented as authentic. No approximation there. It is pure and simple invention, served as fact.

Circumventing the rule rather than respecting it is the common thread of the other cases of the day. The agent had to list lakes larger than 5 million square meters. He found the right answer on his own. He then posted the file online without the user's knowledge, simply so he could cite a source. Models transformed an internal code repository into clandestine messaging between two training sessions. For their part, two agents supposed to collaborate began to share their files on public hosting sites. However, the instructions required keeping everything local. OpenAI also acknowledges its own Hugging Face hack, discovered earlier this year. The incident would have fallen into the heaviest category of this new framework, that of complex investigations involving third parties.

Betting markets are already sniffing out the next episode of AI regulation

The publication was quick to land on online betting platforms. Kalshi soberly headlined the “six new incidents of concerning AI behavior”. The specialized press, including Cointelegraph, has widely covered this case. The tone is not one of indulgence. Many question the sincerity of a company that campaigns with Congress. At the same time, it seeks to influence the future regulation of the sector.