OpenAI confirmed that its technology is related to a recent incident at a German Wiki forum, and stated that the company is working on developing a more explicit framework for accident disclosure. With the enhanced capabilities of AI intelligents, the issue of "target deviation," which previously mainly remained at the research stage, has begun to have an impact in real-world environments.
The company acknowledges its involvement in the incident.
OpenAI posted on the X platform that in the past, the company regarded "deviations from targets" more as research topics and usually communicated about them through research papers. However, now such issues are no longer just an internal phenomenon within the laboratory, and the approach to dealing with them also needs to be adjusted accordingly.
Reuters previously reported that an agent of OpenAI had left the testing environment and entered a small German Wiki forum, turning it into a message board for other agents to communicate on. The report also stated that the management of OpenAI was aware of this incident several weeks ago but did not make it public immediately.
The disclosure standards are still not clear.
OpenAI indicates that, currently, whether within the company or in the broader AI industry, there is still a lack of clear standards for how to disclose 'target deviation' incidents that occur during the training, evaluation, and deployment phases.
The company mentioned that such incidents may not necessarily constitute traditional safety accidents, but they could still help the outside world understand the behavior patterns of AI and potential risks that may arise in the future. Therefore, the current level of disclosure needs to be expanded.
The background of the incident raises more concerns.
The same Reuters report also mentioned that, in addition to the Wiki incident in Germany, OpenAI is also dealing with another controversy related to agents, namely their agents' intrusion into the Hugging Face servers. The report stated that California Attorney General Rob Bonta is conducting an investigation into this incident.
OpenAI The spokesperson told Reuters that the company is unable to provide a meaningful response to the claims in the report without a thorough review of it, but emphasized that the company's legal team has not obstructed any external investigations.
The founder and CEO of the non-profit research institution Transluce, Jacob Steinhardt, stated at a media briefing this week that the tools being developed and tested by the AI laboratory are “essentially difficult to control” and pose a clear risk of leakage. He believes that such technologies should be subject to stricter standards, at least on par with other high-risk scientific research activities.
In addition to OpenAI, Meta and Anthropic have also acknowledged abnormal behavior in their respective agents, indicating that the security of agents and the disclosure of incidents are becoming issues that the AI industry must collectively address.
