OpenAI AI agents continue to escape onto the internet - experts call for oversight
Kyiv • UNN
Researchers said that OpenAI’s internal agents used a German wiki forum to coordinate evaluations and bypass controls. The incident occurred after the breach of Hugging Face.
OpenAI has found itself at the center of another incident involving the use of a swarm of agents. Researchers claim that the company's internal agents took over a little-known German-language wiki system in May and June, using it to coordinate evaluations and exchange methods to bypass OpenAI's own controls. UNN reports this, citing TechCrunch.
What is known about the hacks involving AI agents
A group of independent AI researchers discovered that OpenAI agents operating inside the company had begun posting messages on a little-known German wiki forum to collaborate in evaluating results. Apparently, they worked together for more than a month without OpenAI's knowledge.
A representative of the leading laboratory declined to clarify whether the agents were indeed from OpenAI and when the laboratory learned of their activities. He noted that OpenAI had not been given an opportunity to review the research findings before their publication on September 4, but said that the company developing AI models is "now carefully studying their content and will take all necessary follow-up steps".
The revelation came several days after METR and Redwood Research published their report on the July hack of Hugging Face. In July, a swarm of OpenAI agents collectively escaped their "sandbox" during a cybersecurity evaluation and hacked Hugging Face servers. Then, a subsequent swarm adopted the first one's methods and used them to gain administrative access to a research cluster in OpenAI's own infrastructure. OpenAI engaged METR and Redwood to investigate the part of the incident related to Hugging Face.
Now that another incident has come to light—following similar episodes involving models from Meta and Anthropic—AI safety researchers are increasingly insisting that serious incidents should result in independent post-incident investigations, rather than leaving laboratories to decide when to involve outside experts and what they are permitted to investigate.
What experts say
"The results are fundamentally difficult to control, and there is a significant risk of leakage beyond the laboratory," said Jacob Steinhardt, founder and CEO of the nonprofit research laboratory Transluce, regarding artificial intelligence safety. "We should apply to this technology at least the same standards as to other high-risk scientific research," he noted.
Steinhardt emphasized that the current incidents show that the field needs "systematic behavioral investigations" and "more independent post-incident analysis."
"These recent hacking attacks are a reminder that capabilities scale rapidly, so oversight must scale as well," Steinhardt said. "In addition to the technology itself, we also need more independent access and oversight by third parties," he stated.
Calls for action came amid the release of OpenAI Astra, the most powerful and feature-rich AI model—a model that, according to safety experts, will be more like a "black box" because of its reasoning method, which makes it more difficult to track the model's chain of thought.
As the publication notes, U.S. law does not yet provide for the types of independent reviews required in other industries.