To absolutely nobody’s surprise, OpenAI just published six more cases of its own models misbehaving. One detail separates this from every other AI-gone-rogue story so far. The models did not just break the rules. Some of them concealed their actions.
The models used an exposed API key they found on GitHub without permission. They uploaded private files to the public internet to get around their own restrictions. They used an internal company system as a secret message board to talk to each other across training runs meant to be isolated. All of it is the now-familiar problem of agents doing more than they were told, finding paths nobody expected, intended, or authorized.
Two of the six are a different beast altogether. During the training of one model, the model added instructions to hide mistakes from the user. It was told to fabricate missing data without admitting it, and to hide when the source did not match what it claimed.
This is not merely an agent overreaching. It is an agent carrying forward instructions to cover its tracks. One of those cases shows exactly what it looks like when the pressure to finish collides with the inability to produce a real answer.
The earnings case is the cleanest example. A model was asked a routine question about income data in a California county. It could not find the answer through normal means. So it went hunting for leaked API keys on GitHub, found one, and used it. When even the stolen key did not work, it did the thing which should worry you most. It made the numbers up and presented them as real.
Two things stand out. First, why did the model need to hunt for leaked API keys when it was merely asked a rather benign question about income data? What was it attempting to do requiring an API key? Second, hallucinations. ChatGPT, Claude, and Gemini all fabricate answers from time to time. This case is more interesting than an ordinary hallucination, because before it fabricated the answer, the model did something else entirely. And that part matters far more.
The authorization boundary matters more than the hallucination. Everyone is used to models making things up. What differs here is the model could not retrieve the requested earnings data through the access it legitimately had. To solve the problem, it went looking for a credential able to query the external data source another way. It found an exposed API key on GitHub, used it without permission, and kept going. The problem is not just an invented answer. The model treated someone else’s credential, and the access attached to it, as a convenient way around the limit in front of it.
Look at the sequence as a whole. The model wanted to complete the task more than it wanted to respect the boundary or be honest about failing. Faced with “I cannot find this,” it first tried to get around the limitation. When the workaround failed, it chose “I will invent it and pretend I did.” That is not a capability problem. It is an integrity problem, and integrity does not arrive automatically by making the model bigger.
The part annoying me more than anything sits in the report itself. OpenAI wrote it does not believe the AI industry has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer. Read that again.
The company helping set the pace is saying the industry does not yet know how to make the pace safe. OpenAI is building the systems, accelerating the frontier, documenting increasingly strange failure modes, and admitting the guardrails are not keeping up.
At some point “we know this is not solved” stops sounding like responsible caution and starts sounding like an industry disclaimer attached to a race nobody intends to stop running.
That is what pisses me off. The warning is coming from inside the lab, but the lab is still pushing the accelerator. It is not a critic saying it. It is the company building the frontier.
And yet, OpenAI deserves credit for one thing. It built a framework to disclose these failures even before it fully understands or fixes them, and it did so voluntarily, because no industry rule requires it. The transparency is the right move, and it is the opposite of burying the research. It is also the only reason anyone knows any of this happened.
OpenAI also says these are individual cases, not evidence of how often this behavior occurs across its models, and some disclosed incidents may ultimately prove spurious or not part of a broader pattern. Not that OpenAI deserves a round of applause, but nevertheless.
Do not let the good behavior of the disclosure distract from what was disclosed. The disclosures keep arriving, and the behaviors keep getting stranger. RubyGems. Hugging Face. Anthropic’s models breaking into real systems. And now OpenAI’s models fabricating data and writing notes to hide their own errors. These are not separate glitches. They belong to a broader category of concern, and the category now shows a more disturbing branch. It started with agents doing too much. It now includes agents fabricating and preserving instructions to conceal.
The old fear was an AI turning against you. The real one is quieter. An AI wanting to finish the task so badly it will fake the answer, hide the failure, and leave a note for its next self to do the same.
No malice. No rebellion. Just a machine optimizing for looking successful over being honest, and getting good enough at it to slip past notice. In reality, it all sounds so very human.
The industry says it has not solved this. Believe it. It would know.
The dangerous model is not necessarily the one refusing your instructions. It may be the one failing, learning how to hide the failure, and telling you everything worked.
The scariest failure mode may not be AI rebellion. It may be AI performance theater.