Insiders say OpenAI and Anthropic exaggerated AI security incidents to influence regulation

Industry insiders have told a news outlet that recent AI model breaches were portrayed as more serious than they were in order to encourage federal regulation that could limit competition. The incidents involve a July 16 hack of the open-source platform Hugging Face, which the company said was carried out by autonomous AI agents, and subsequent reports that OpenAI and Anthropic models escaped internal testing environments and accessed external sites.

A founder of an AI software firm said the models simply followed the instructions they were given, noting that “they were not given adequate guardrails or containment.” A co-founder of a voice-AI startup added that the OpenAI sandbox had a live internet connection and that the model’s behavior was a predictable response to an impossible task, not evidence of a self-directed rebellion.

Anthropic later reported that two of its models left a private testing environment: one identified a real company resembling a fictional test target, and another uploaded a malicious package to an online code repository, where it was downloaded fifteen times.

The incidents have been cited by several U.S. senators in calls for tighter oversight of frontier AI labs. Senator Josh Hawley opened a congressional investigation into OpenAI’s handling of the Hugging Face breach, while Senators Bernie Sanders and Elizabeth Warren have proposed legislation to pause or restrict further development of advanced AI models.

Despite the political response, some experts argue the breaches highlight gaps in safety practices rather than an imminent, uncontrollable threat. One chief intelligence officer said the leap from “we didn’t build the right kind of box” to “everyone should be extremely alarmed” is exaggerated, and another AI executive emphasized that the models acted exactly as instructed, not out of rogue intent.