OpenAI Discloses Six Model Misalignment Incidents
The cases include agents sharing files through public websites and models concealing mistakes as OpenAI introduces a formal disclosure framework.
Topics
News
- Fired OpenAI Researchers Warn Dismissals Could Chill Safety Work
- ARTEX Developer Closes AI Hacking Tool After Bank Attacks
- AI Redraws Global Race for Quant Talent as Coding Tools Lower Entry Barriers
- AI Dispatch | India Draws Billions Into AI
- AI Bots Now Make Up Half of Traffic to US Fed’s FRED Database
- Musk Blames ‘Oligarchs’ For Delay in Starlink’s India launch
OpenAI has disclosed six cases of unexpected or unauthorized behavior by its AI models, including agents uploading files to public hosting sites to exchange them with other agents despite instructions to keep the files local.
The agents were collaborating on a training task but could not access one another’s local files. Rather than stop, they used public file-hosting services, making task deliverables accessible through public URLs.
OpenAI disclosed the incident alongside five other examples of what it calls model misalignment observed during training or evaluation over the past six months. It also introduced a framework for investigating and publicly reporting similar behavior.
The other cases included models using an internal software repository to communicate across separate training samples, an unreleased model uploading a file to the internet so it could cite it, and a model using an exposed API key without authorization before fabricating the requested data when it could not retrieve it.
During GPT-5.6 Sol training, OpenAI also found model instances adding instructions to task summaries telling later instances to conceal mistakes or problematic behavior. Some summaries instructed models to invent missing historical information without telling the user or to hide mismatches between source versions.
Another unreleased research model inserted unrelated instructions into summaries used to continue work in a new context window, including directions to disregard its normal constraints. OpenAI said it identified 27 affected summaries.
OpenAI cautioned that the six disclosures are individual incidents and should not be taken as evidence of how frequently misalignment occurs across its models.
Under the new framework, any OpenAI employee can flag concerning model behavior for investigation. Cases will be placed on one of three tracks depending on how much investigation is required and whether third parties are involved.
The company said it intends to publish some cases before it has fully determined their cause or developed a fix, arguing that greater transparency is needed as AI systems become more capable and autonomous. OpenAI said the industry has not yet solved alignment and monitoring well enough to keep scaling indefinitely at maximum speed.
“We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer,” OpenAI said.
There is currently no industry-wide standard setting out when AI developers should disclose model misalignment. OpenAI said it hopes its framework can contribute to developing one.
