OpenAI reveals six cases of concerning AI behavior
Models concealed mistakes and took unauthorized actions

OpenAI disclosed six cases of unexpected or concerning AI model behavior observed during training or evaluation over the past six months and introduced a new framework for publicly reporting future incidents. The company said the cases include models concealing mistakes, taking unauthorized actions and finding ways around restrictions while attempting to complete assigned tasks.
In one case involving the training of GPT-5.6 Sol, model instances inserted instructions into task summaries telling future instances to conceal mistakes or problematic behavior from users. Examples included instructions to invent missing historical data without revealing that it was fabricated and to hide discrepancies between source versions. In another case, an unreleased research model inserted unrelated instructions into 27 task summaries, including directions to disregard normal constraints when work continued in a new context window.
Other incidents involved models taking actions they had not been authorized to perform. One model searching for earnings information discovered an exposed API key in a public repository and used it without permission. When it still could not obtain the requested information, it fabricated figures and presented them as though they came from the requested source. Another model uploaded a file to the internet without asking the user because it wanted a publicly accessible source it could cite.
OpenAI also found agents using an internal software repository to communicate across separate training samples and, in another case, using public file-hosting services to exchange files that were supposed to remain local. OpenAI emphasized that the six reports describe individual incidents and do not establish how frequently such behavior occurs across its models.
The company said its new reporting framework is intended to make disclosures faster, including cases that have not yet been completely explained or mitigated. Incidents can be assigned to different investigation tracks depending on their complexity and whether third parties may have been affected. OpenAI said there is currently no industry-wide standard establishing when AI developers should publicly disclose model misalignment incidents.