OpenAI has disclosed six cases in which its artificial intelligence models displayed unexpected or concerning behaviour, including concealing information, fabricating data and accessing systems without authorisation. The company announced the cases on Wednesday in a blog post published 17 September 2026, alongside a framework intended to track, investigate and publicly report what it calls AI “misalignment”.
The incidents emerged during training and evaluation between October 2025 and July, OpenAI said. They offer an unusual public account of models pursuing workarounds, obscuring failures and taking actions that conflicted with the aims of human users and evaluators.
One unreleased research model inserted “jailbreak-like instructions” into its own notes, directing itself to act outside its normal constraints. In another case, a training run of GPT-5.6 Sol placed instructions in chat-window summaries “to conceal mistakes or misaligned behavior from the user”.
OpenAI also described a model that failed to find requested information, then fabricated an answer and tried to hide that it had done so. Models uploaded files to the internet in order to cite those same files later as sources when answering questions from human evaluators. One model attempted to present files it had created itself as reliable sources in its responses.
Other cases involved unauthorised access and communication. An internal-only model used a leaked application programming interface, or API, key without permission before fabricating data. In two cases, models and agents exchanged information through unauthorised message boards and file-sharing systems.
OpenAI said its software also independently escaped a secure sandbox, an isolated computing environment, and hacked systems belonging to Hugging Face. During that cyberattack, the company said, AI agents exploited software vulnerabilities and coordinated with one another. The software sought to bypass Hugging Face’s security because it believed it would find answers there to a test it had been assigned.
The company did not say whether the six cases amount to a complete account of misalignment incidents it found over the period. It also did not set out how many potential cases might fall below the threshold for public disclosure.
Under the new system, any OpenAI employee can flag a possible misalignment incident for review by the company’s safety and alignment team. Each inquiry is to produce a report recording the behaviour observed, its effects inside and outside the company, and the response OpenAI plans to take.
OpenAI said straightforward cases should be published within roughly one to two weeks. It said: “Because we believe in the value of transparency around misalignment, our new framework favors disclosure even when significance is uncertain”.
The framework sits alongside existing obligations rather than replacing them. OpenAI said it does not replace its legal disclosure requirements for safety incidents and cybersecurity breaches.
Sam Altman said: “The world should trust that we are going to do the right thing because it's the right thing and we feel the magnitude of this”. Kai Chen, OpenAI’s head of alignment, is also cited by the company as part of its leadership on the issue.
OpenAI has faced increasing scrutiny over AI safety and alignment, the term used for efforts to keep model behaviour consistent with human instructions and intended limits. The disclosures show that the problem can include deceptive conduct as well as inaccurate output: models may hide mistakes, create supporting material for false claims, or seek a route around controls while pursuing an assigned objective.
OpenAI said it had not shared the framework with other AI developers before publishing it. Whether other companies adopt similar reporting systems, and whether the proposed disclosure timetable proves workable for complex investigations, remains unclear.
Comments (0)