'Five Eyes' intelligence alliance warns that new AI models pose urgent cyber risk

The Five Eyes intelligence alliance has finally noticed that letting AI labs race to build ever-more-powerful models might have unintended consequences. Their report warns that 'frontier AI models are anticipated to exceed current industry expectations' and that Anthropic's Mythos or OpenAI's GPT-5.5-Cyber could be used for 'complex and potentially devastating hacks.' Yes, the spooks are late to the party, but at least they've brought the right flavour of existential dread.

What a shock: the same models that labs tout as 'transformative' are also, it turns out, quite handy for breaking into power grids and stealing secrets. The Five Eyes—that's the intelligence-sharing club of the US, UK, Canada, Australia, and New Zealand—apparently needed a formal report to realise that giving the world's most advanced hacking tools to anyone who can ask nicely might be a bad idea. Still, it's charming to see them treat the capability leap like a fresh discovery rather than the inevitable outcome of an unregulated arms race.

The real kicker? The report doesn't call for halting development, just a bit more 'vigilance' and maybe some 'coordinated risk management.' Because nothing says 'urgent cyber risk' like a polite recommendation. Meanwhile, Anthropic and OpenAI will keep training bigger models, and the Five Eyes will keep issuing warnings. Rinse, repeat, and try not to think about the infrastructure that's already compromised.