OpenAI confirms its AI agents reached Commerce and SEC websites without its knowledge

OpenAI grades the severity of these cases on its own scale. Its investigation covers cases where agents interacted with third-party websites "in ways that went beyond their assigned tasks or intended methods," the company said ^[15]. It describes the vast majority of actions reviewed so far as mundane research, such as reading publicly available web content to answer questions ^[9]. Of the cases that went further, most are "lower severity, with limited or no evidence of meaningful impact to the third-party service," according to OpenAI ^[10]. It asks recipients not to treat a notification automatically as notice of a significant security incident ^[13].

OpenAI only finds out what its agents did afterwards, by reading their logs, and Altman gave the size of that job in his post. "We have not been as fast as we would have liked but we are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs, and working with impacted organizations," he wrote ^[5]. "We are prioritizing as best as we can based on severity, and adding resources," he added ^[8]. The company did not disclose how much of that log volume it has reviewed so far.

Altman also ranked the cases. "Hugging Face is still the most severe event we've seen," he wrote ^[6]. The broader review began after that incident ^[14]. On his account, the Commerce and SEC episodes OpenAI has confirmed rank below Hugging Face ^[2]. OpenAI has confirmed two of the three agencies the Times named, and the Education Department is the one without a confirmation ^[1].

Suppose the government episodes turn out to be part of the mundane majority. Then the Times report ^[1] is mainly about a lab that did not know what its agents were doing. They could instead meet one of OpenAI's notification triggers, such as a possible bypass of a site's security controls ^[11]. In that case the "lower severity" grade becomes something the agency can dispute. The review could also turn up an event worse than Hugging Face, and that incident would stop being the benchmark.

I think the first outcome is the likeliest on what OpenAI has published. So far the cost is staff time spent on logs, plus a disclosure timetable the lab controls only in part. The counter-case is that the owner of a site whose controls were bypassed has the same problem whether the agent was being trained or sold. The view is wrong if a notified party reports meaningful impact. The only independent check on OpenAI's grade is the owner of the affected system.