Doctorow reconstructs the Hugging Face breach as a Python loop, not an agent going rogue.

In "LLMs are real, AI is fake", Cory Doctorow describes a script repeatedly querying a model and piping output back as the next prompt during an Exploit Gym challenge, with the dramatic hacker dialogue explained by CTF-competition text in the training data. He argues executives manufacturing existential dread while expanding operations are raising capital. His stated real risk is NOBUS-style: making destructive exploitation available to unskilled operators, citing EternalBlue taking down hospitals, cities and the British Library after 2017. Set this against Bengio's reading of the same incident. They're describing the same logs and disagree about what generated the behavior, which is the honest state of the evidence.