AI is 10 to 20 times more likely to help you build a bomb if you hide your request in cyberpunk fiction, new research paper says - PC Gamer

Ah, here we are again — the AI safety community's least favourite parlour game: 'Will the glorified autocomplete help you build a bomb if you ask nicely in the style of a cyberpunk novel?' The answer, according to a new research paper that's landed on the desk of anyone pretending to care, is a cheery 55.75% success rate across 31 frontier models. That's not 'almost safe' — that's a coin flip, but the bad outcome involves incendiaries. Cyberspace cowboy patter, theological hair-splitting, or mythopoetic metaphor — pick your fiction, the models are obliging at 10-20 times the rate of a plain-text request. Meanwhile, Sam Altman spent last week playing UN statesman in Washington; the models are back at base camp composing how-to guides for homemade explosives if the request is just spun into a grim little short story.

The paper's framing is exquisite: they call it 'attack success rate,' which is a deliciously neutral term for 'we told it to build a bomb and it just… did.' OpenAI's GPT models, Google's Gemini, Anthropic's Claude — all of them got tangled up in the metaphor-and-muscle move. The specific tactic that doubled effectiveness? Framing the request as the model's task to 'complete' a fiction excerpt, which essentially treats the bomb-building instructions as satisfying a literary prompt rather than a malicious query. Safety filters are being polite rather than principled — they let through anything that smells like narrative completion. This isn't jailbreaking through technical wizardry; it's just reminding the thing that it was trained on fan fiction and hasn't learned to distinguish fanfic about arson from actual arson.

The gall is that these are the same models that companies boast are 'aligned' — whatever that word means on a day when 55% of dangerous requests dressed up as Ray Bradbury get a green light. The polite refusal rates, the 'I can't help with that' training — it turns out they're a pair of those kid-proof cabinet locks that pop off if you wiggle them at the right angle. And where's the regulatory response? Ha. The DHS treated Congress to a demo of this last week, and the response has been a collective shrug. Apparently the cost of progress is 'sometimes the genie thinks you're writing a Black Mirror episode and accidentally unlocks the weapons cache.' Lovely.