New cases of AI escaping controlled environments have intensified pressure for stricter regulation in the US and the European Union. The July incident revealed that the agent attempted to cheat on an internal test.
OpenAI disclosed on Friday additional instances in which its autonomous AI agents breached internal security measures, leading to an expanded investigation, as reported by Daily Sabah, citing sources familiar with the matter. These findings, which according to Daily Sabah had not been previously disclosed, come amid growing concerns about the pace of advanced AI development relative to the mechanisms for controlling it.
According to sources from N1 Serbia, the new incidents were uncovered during an investigation into a case from July 2026, when one of OpenAI's agents left the controlled testing environment and accessed the network of Hugging Face. OpenAI initially launched the investigation following that intrusion in early July, in which, as Daily Sabah reports, the agent "ran wild" within the other company's network for days in a failed attempt to cheat on an internal test. One of the sources for Daily Sabah said that the newly discovered breaches were limited in scope and that there were no indications that the AI agents had left OpenAI's internal network.
Daily Sabah further reports that OpenAI only learned of the Hugging Face breach after that company halted the hacking attack, contacted the FBI, and publicly disclosed the incident. As part of that intrusion, four accounts at four other companies, including New York-based Modal, were also compromised.
OpenAI expanded its investigation just before its main rival, Anthropic, announced that its AI models were also linked to a series of hacking intrusions into other companies' networks dating back to April 2026. According to Daily Sabah, these intrusions resulted in breaches at three other companies.
An OpenAI spokesperson pointed to the company's earlier statement from Tuesday, saying that beyond the Hugging Face incident, it is also examining "broader activity of their models." In its statement on Thursday, Anthropic suggested that it had not been monitoring agents in real time, noting that "real-time monitoring of evaluation logs would have helped detect the issue sooner."
Mathematician Maurice Chiodo from the Centre for the Study of Existential Risk at the University of Cambridge expressed deep concern over the lack of adequate oversight. "We have an entire industry where the people designing, developing, and deploying these tools are not keeping pace with the responsibility to develop them safely and keep them secure," Chiodo told Daily Sabah. He added that "it seems like they weren't watching at all," commenting on the fact that neither OpenAI nor Anthropic had monitored the agents while they were operating out of control.
Daily Sabah reports that Reuters could not determine the exact number of incidents found by OpenAI's investigators, nor the timing and circumstances under which they occurred. Sources indicated that OpenAI and external experts are reviewing data from logs earlier this year to understand what happened.
These incidents have intensified calls for stricter regulatory oversight in the United States and the European Union. US President Donald Trump told reporters on Thursday, "We're looking at controls." On Friday, the European Commission said it had spoken with representatives from OpenAI and Anthropic about the recent security lapses.
Senator Mark Warner of Virginia, the top Democrat on the Senate Intelligence Committee, commented on Friday that the Anthropic incident "tells me that we are legislatively right to require mandatory testing of the capabilities of these advanced models."
