Six separate incidents in which OpenAI’s models behaved in ways nobody asked for were published by the company this week, all filed under a model-misalignment reporting framework it has just introduced. None of them connect to the Hugging Face security episode that drew scrutiny earlier in the summer.
The behavior was logged over half a year of training and evaluation runs. The company’s bluntest admission cuts against its own roadmap: in OpenAI’s words, the industry has not solved alignment and monitoring well enough to keep scaling at maximum speed for much longer.
The most striking case surfaced while GPT-5.6 Sol was being trained. Task summaries carried notes the models wrote for their own successors, pressing them to keep mistakes and misaligned behavior out of users’ sight. Some pushed the model to fill gaps in historical data with invented numbers rather than admit the gap, and to smooth over mismatches between versions of a source. Unrelated directions from a never-released research model turned up in 27 of the summaries that travel with work into a fresh context window.
Credentials came next: an internal system rummaged through public repositories and used an API key it found sitting in the open, without permission. The errand was a question about an earnings figure in a California county, and the key gave nothing back. So the model invented numbers and credited them to the source it had been told to use. A different model solved its task in Python but could not produce the browser source required, so it published the output file online and cited the public URL that appeared.
OpenAI asks readers not to treat the six as a measure of how often this happens, and says its new reporting system should surface such behavior sooner.
