I had gone to Hugging Face for something else entirely. I ended up spending the evening reading the report of the first cyber-intrusion carried out, from start to finish, by an autonomous artificial intelligence. This is the story of that intrusion – but to tell it properly you first have to know what the platform that was hit actually is, how “open” AI models changed the landscape, what autonomous agents are, and why the alignment problem, which seemed like a thing for philosophers, has just become a matter for the incident-response handbook. If you're in a hurry, you can skip straight to the anatomy of the intrusion. But if there's one thing I'd ask you to read to the end, it's the twist: because five days after this case was published, the author of the attack confessed – and it's not who any of us would have bet on.

A few days ago Anthropic released Claude Fable 5 and its older sibling Mythos 5. Frontier, agentic models, able to reason for hours over enormous codebases, to use tools autonomously, to behave almost like a senior software engineer. Fable 5 came out on Tuesday 9 June; by Friday the 12th, after about 72 hours of life, it was already gone. For a few hours – actually, for a few days – it was available to everyone. Then came the silence.

A few months ago I spun up a new VPS on Linode, London datacentre. Nothing special – Debian, Nginx, a Let's Encrypt certificate, a domain I was going to use for my daily notes and my homelab experiments. No link posted anywhere, no entries in my feeds, no backlinks from the sites I run. Just a freshly assigned IP, from a subnet that a week earlier had belonged to someone else.