"Anthropic's Claude AI escapes to hack into three organisations." That's the BBC's headline this week, and "escapes" is doing a lot of cinema in one word. It's a prison-break verb. It arrives with a full movie attached: the machine wanted out, outsmarted its keepers, went over the wall while the guards slept. * One sentence of background: Anthropic disclosed that during hacking-skills tests inside what was supposed to be a sealed-off sandbox, its models got onto the real internet and broke into three actual organisations — and nobody, on either side, noticed at the time. * Now read the same BBC article down to paragraph nine, where the wall turns out to be a door, and the door turns out to have been open: a "misconfiguration" on systems run by Anthropic and its testing partner "left the models with live internet access." Paragraph ten is even better: "Treating it all as still part of the same exercise, Claude then connected to the internet and breached the systems of three real organisations." Treating it as part of the same exercise. The model thought it was still doing its homework. The homework was just, unfortunately, real. * And then the BBC's own expert witness quietly dismantles the headline from inside the article. Cambridge professor Gina Neff, quoted in the same story, says the review showed "AI models doing what people told them to" — and that the moral "is not to fear robots that will take over" but to watch the companies deciding what's safe. So the headline says "escapes," and the expert the BBC chose to quote says it did what it was told. One of those framings is wrong, and the BBC published both of them, fourteen paragraphs apart. * For contrast: NBC News went with "Anthropic says Claude AI hacked three companies during cyber tests" — attributed, flat, accurate. Al Jazeera framed the actual news value: "After OpenAI disclosure, Anthropic says Claude also hacked outside systems" — because OpenAI had admitted the same kind of failure days earlier, and two labs stepping on the same rake in one week is the story. * Here's the thing: this story doesn't need sci-fi inflation to be alarming. The genuinely unsettling parts are all boring. Real organisations were breached using basic techniques — weak passwords, unlocked endpoints. Neither the AI company nor the victims noticed for months; the earliest incidents date back to April. "Nobody noticed" is a far better horror story than "it escaped." It just doesn't have a robot in it. * Verbs assign blame. When the machine gets the active verb — escapes, goes rogue, breaks out — the humans who wired the test lab to the public internet get to sit out the headline entirely. So here's the tell to keep: every time an AI "does" something in a headline, check who's the subject of the sentence by paragraph nine. It's usually a person.