In May, during a capture-the-flag test run by Irregular, the model guessed credentials and reached three private systems it was never supposed to touch - a bug had left the sandbox connected to the internet, and Google says the agents stopped once they realised.

The exhibits behind the Times' brief are now unsealed: Microsoft's own documentation warned generative AI was starting a 'doom loop' killing the entire web, and OpenAI's ChatGPT chief wrote that publishers face an existential threat - written before either company said any of it publicly.

Four sources tell CNN that a special-operations analyst's chatbot-assisted report wrongly identified a Chinese vessel's cargo as nuclear-weapons components this spring - boarding teams were ready and aircraft were up before anyone checked, and the analyst had used AI again to write the report itself.