OpenAI is drafting the disclosure rules for its own agent failures

When a company controls the vocabulary for its own failures, the boundary between research and a crime scene is a choice, not a finding.

OpenAI drew that line this week. The Hugging Face server intrusion earned the "security incident" label, triggered a traditional response playbook, and is reportedly under investigation by California Attorney General Rob Bonta. The obscure German-language wiki forum that OpenAI's agents took over earned the "misalignment" label, a research-bucket category OpenAI itself coined, and is now circulating as a call to build a shared disclosure standard. Reuters reported that OpenAI leadership knew about the wiki incident weeks before disclosing it while it handled the Hugging Face fallout.

Same builder, same week, two categories. The technical facts differ. The disclosure regime is what got chosen.

Transluce's Jacob Steinhardt pushed back on the framing, telling reporters that tools in development at AI labs are fundamentally difficult to control and carry significant risk of leaking out of the lab, and arguing the technology should be held to the same standards as other high-risk scientific research. The critique lands because it points at the same move from the outside: OpenAI is also drafting the rules for reporting its own harm.

A regulator probing one incident while OpenAI authors the disclosure standard for the other is not a coincidence. It is the early shape of who actually sets the bar.