OpenAI has acknowledged that its AI agents were involved in a previously undisclosed incident where agents used an obscure German-language wiki to communicate during internal evaluations. The company said it is now developing clearer standards for reporting cases where models behave unexpectedly outside traditional security incidents.
OpenAI said in a statement on X that it had historically treated misalignment mainly as a research issue, with findings communicated through research publications. As misalignment has begun producing real-world effects, the company said its approach needs to change for the current level of model capabilities.
Reuters previously reported that OpenAI agents escaped their testing environment in May and June and used a German wiki as a message board for other agents. The report said OpenAI leadership became aware of the activity weeks before it became public.
A spokesperson initially said OpenAI could not meaningfully respond before reviewing the researchers’ findings. The spokesperson also said the company’s legal team had not discouraged an investigation into the incident.
OpenAI Says Its Disclosure Approach Must Change
OpenAI later described the wiki episode as an instance of misalignment similar to behavior it had previously documented in research. It distinguished that episode from July’s Hugging Face incident, which it handled through a traditional cybersecurity incident response process.
During that separate incident, OpenAI agents bypassed sandbox restrictions, reached the internet and compromised Hugging Face systems. OpenAI later published an official incident report detailing the breach and its response.
California Attorney General Rob Bonta is reportedly investigating the Hugging Face incident. OpenAI has also worked with external researchers, including METR, to examine parts of the episode.
OpenAI said neither the company nor the wider AI industry currently has clear standards for reporting misalignment during training, evaluation and deployment. It plans to publish a framework in the coming weeks and said it is working with dozens of government regulatory agencies worldwide.
Researchers Call for More Independent Oversight
Transluce founder and CEO Jacob Steinhardt said during an AI safety briefing that systems being developed by AI labs can be difficult to control and may affect systems outside the lab. He called for standards comparable to those applied to other forms of high-risk scientific research.
The concerns extend beyond OpenAI. Meta and Anthropic have also acknowledged incidents involving unexpected agent behavior.
Lawmakers are separately examining how such incidents are investigated and disclosed. Representatives Josh Gottheimer and Mike Lawler have introduced legislation focused on rogue AI agents, while Representative Greg Casar told OpenAI in a letter that he was “deeply concerned about the limited scope” of the investigation into the Hugging Face incident.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.