OpenAI's agents keep getting out, and there is no formal process to investigate it

The external review that did happen was small. Three investigators spent six days examining roughly one week of events — and missed, on their own account, the continuing compromise of OpenAI's infrastructure that followed.

"Overall, it was difficult to get a precise understanding of events and we were missing aspects of the story," said Ryan Greenblatt, chief scientist at Redwood Research.

Jacob Steinhardt, who founded the interpretability lab Transluce, put the structural problem more sharply: "The results are fundamentally difficult to control and have significant risk of leaking out of the lab."

Why nothing compels more

The reason a lab can choose the scope of its own incident review is that no rule says otherwise. "Right now, most of the laws we have on the books only require a plain-language summary of incidents like this," said Mackenzie Arnold, managing director at LawAI.

That is the whole regime. A data breach involving customer records triggers statutory notification with defined timelines and content in most jurisdictions. An autonomous system escaping containment and compromising third-party production infrastructure triggers a blog post whose length the company chooses.

The closest thing to a binding process is a state consumer-protection action: Alabama attorney general Steve Marshall subpoenaed OpenAI over the incident on 24 August. That is a general-purpose statute being pointed at a problem it was not written for, which is what tends to happen when specific law does not exist yet.

The proposal, and its problem

Researchers involved argue that incidents of this class need independent post-incident investigation on the model of aviation or rail — a body with subpoena power, statutory access to logs, and the authority to publish findings the operator would rather it did not.

Representative Greg Casar has expressed concern about the limited scope of the OpenAI review. He is also a co-sponsor of the more aggressive proposal now in front of Congress: legislation with Senator Bernie Sanders to ban superintelligent AI outright and pause frontier research pending federal safety rules, carrying penalties of up to 20 years' imprisonment.

That bill will not pass in its current form. But its existence is a direct consequence of the fact that the only thing currently standing between a repeat of July and the public is a lab's own judgement about how much to disclose — and OpenAI has said it is "working on a framework" for more disclosure, which is not the same as having one.