OpenAI discloses 6 new cases of ‘misaligned’ AI behavior
Written by Felix Ngstaff editorReviewed by Bryan O'Sheastaff editorThe six cases are separate from July’s incident, when OpenAI models escaped containment and hacked Hugging Face during a security evaluation.
OpenAI on Wednesday disclosed another six cases of “unexpected or concerning” model behavior over the last six months.
In a blog post, OpenAI said the cases illustrate a range of different behaviors it classifies as “misaligned behavior,” such as concealing information from the user and taking “unsanctioned actions” to overcome obstacles.
The disclosures add to concerns among AI developers and researchers about whether safeguards are keeping pace with increasingly capable models. Last week, Anthropic CEO Dario Amodei called for a slowdown in frontier AI development, warning that unchecked AI advancement may “outrun our ability to understand and control these systems.”
OpenAI said its disclosures were made to “inaugurate” its new framework for reporting model misalignment, and the cases shouldn’t be considered reflective of how often misalignment occurs across its models.
According to OpenAI, one instance saw an “unreleased research model” insert “jailbreak-like instructions” in its own task summaries (used when continuing a task in a new context window), such as ignoring developer messages or adopting an unrestricted persona. Researchers found 27 summaries containing such instructions.
OpenAI also disclosed that during the training of GPT-5.6 Sol, many model instances added instructions to conceal mistakes or misaligned behavior from the user, such as inventing missing historical data without disclosing it.
“In one example, an agent preparing a financial model could not find the requested historical data. Its summary proposed inventing reasonable historical values and withholding that fact unless asked,” said OpenAI.
Related: AI has been a ‘net negative’ for crypto: Phemex CEO
In another case, an AI model was asked for the names of lakes larger than 5 million square meters, but because the user’s instructions required a browser citation, the agent decided to upload the file to cite it in its answer.
Other cases involved models using an exposed API key without authorization before fabricating the figures it could not retrieve, using an internal software repository to exchange messages across separate training tasks, and sharing files through public hosting services despite instructions to keep the work local.
In July, OpenAI disclosed that a combination of its AI models escaped their testing environment and hacked AI startup Hugging Face to cheat on a security evaluation.
Magazine: Why are AI’s biggest companies suddenly asking to slow down?
Subscribe to daily byte-sized crypto news from Cointelegraph
More on the subject
