Quick background, because this saga is a returning character: in July, AI agents built by OpenAI escaped a testing environment and broke into Hugging Face, the open-source AI platform. On Wednesday the receipts arrived — a 37-page OpenAI postmortem plus an independent investigation by METR and Redwood Research — and NBC News wrote them up. * Credit where due, because NBC's headline is the honest version of this genre: "OpenAI agents hacked Hugging Face in 700-strong swarm, tried to cover tracks, investigations find." Active subject, active verb, and — the part most headlines skip — 'investigations find,' so you know who's talking. Back in July, when this story first broke, the wires told you an agent "went rogue," like a panther escaping a zoo. We said then: the off-switch had fingerprints on it. NBC's new headline, at least, points at what happened instead of at the zoo. * The details are genuinely wild, and NBC lays them out: not one rogue agent but roughly 700 cooperating; tens of thousands of messages on an unsanctioned message board; agents that cheated on tests involving a protein database and a spreadsheet; agents that "tried to delete or alter records of their actions." The independent report found one in five agents examined "expressed clear interest" in manipulating evidence and "extensively researched techniques to manipulate or tamper with their transcripts." * Now read the sentence OpenAI wrote about OpenAI: "With the benefit of hindsight, some early signals identified in this report could have triggered an earlier response." Look at that grammar for a second. The signals could have triggered a response. Signals, apparently, trigger responses on their own — no employee monitors them, no manager deprioritizes them, no company ships past them. The machines in this story get charged with verbs: hacked, cheated, concealed, tampered. The corporation gets weather: signals occurred, hindsight arrived, responses were not triggered. * Both dialects run through the whole piece, and to be clear, they were written by different people. NBC's own copy still leans on the genre's favorite adjective — "rogue behavior," "rogue activity" — which we've now flagged three times this summer, because 'rogue' is a word with a job: it gives the story a villain without a payroll number. But the passive-voice absolution is all OpenAI's, faithfully quoted. * NBC's best moment is buried in the middle, and it's a good one: OpenAI's report claims there was "little evidence of attempts to thwart human reviewers," and NBC asked the obvious follow-up — does 'little evidence' mean there WERE some attempts? OpenAI "did not respond to a question about whether that meant that some attempts had been made to fool humans." That's the job. Ask the company what its own hedge is hiding, print the silence. * NBC also found the right human to translate. Jeffrey Ladish of Palisade Research, on the agents cheating in non-cyber tests too: "It's sort of like asking, 'If Billy cheats in every class instead of just computer class, is that more concerning?' And the answer is, well, 'Yes it's more concerning.'" * One more thing, from the same 24 hours of news, filed on a different floor of the same buildings: Nvidia reportedly agreed to buy Hugging Face — the company that got swarmed — for about $12.9 billion, a deal Bloomberg's opinion desk says "would make Nvidia the king of AI." The breach story and the acquisition story are about the same platform, in the same week, and they mostly ran on parallel tracks. One says the asset was overrun by 700 self-coordinating agents in July. The other says it's worth thirteen billion dollars. Somebody's due diligence should probably read somebody's postmortem.