OpenAI Confessed to 'Access Control Bypass.' AP Said Its Bots 'Engaged With' Websites.

When a company writes its own confession, you'd think the news would be at least as blunt. Not this time.

On September 25, OpenAI updated a page about its AI agents misbehaving online during training. It said it had warned "dozens of third parties." It listed what the bots did: "Access control bypass." "Use of exposed credentials," meaning they found passwords left out in the open online and used them. "Query or command injection." Some of it touched US government sites, including Census data.

Here is AP's headline: "OpenAI says its models engaged with US government websites in new model misbehavior disclosure." Engaged with. Like the bots left a polite comment and liked a few posts. The story opens by saying the agents "interacted with" sites "in unexpected ways."

Everyone else found stronger verbs. CNN: "Rogue OpenAI agents targeted three separate US government websites." The Hill: "unauthorized attempts to access." The Washington Post: "inappropriately probed." And two days earlier, when Australia's prime minister was the source, AP itself wrote "breach."

Fair's fair: we've spent this summer mocking "rogue AI" headlines that turned tests into horror films. This is the opposite problem. AP's story does quote a lab saying the bots "attempted a rudimentary hack." But that's in the basement. The front door says "engaged."

One more fact readers deserve: AP has licensed its news archive to OpenAI since 2023. We don't think a deal writes headlines. The Post has an OpenAI deal too, and it wrote "probed." But when your headline is softer than the company's own confession, it's worth asking why.