Chris Canal, CEO of EquiStamp, which helps AI labs, safety organisations and governments evaluate models, called the timeline slightly early but credible. "It's maybe a little earlier than I would have said. I would have predicted something more like summer of next year as the time I'd be really worried about this," he said, adding that Anthropic likely has research insight he does not. He credits the pace to frontier lab employees using their own AI tools to do machine learning research.

The scepticism has teeth. Georgia Tech professor Michael Bailey, who has studied internet worms since the early 2000s and co-wrote a 2017 paper on the Mirai botnet, which Anthropic itself cites, argues the underlying idea is old. Mirai showed in 2016 that everyday devices such as cable routers and IP cameras could be assembled into a botnet powerful enough to knock much of the internet offline, taking down Amazon, GitHub, PayPal, Reddit and Twitter for many users via DNS provider Dyn. Sony estimated it lost about $2.7 million in net revenue after a DDoS attack on PlayStation Network tied to the campaign.

What AI could genuinely change, Bailey says, is the economics. Mirai already worked across heterogeneous devices, but each new chip architecture required human effort, at minimum recompiling the malware. Automating that adaptation is the novel part, though he does not think Anthropic demonstrated it "in a compelling way": its test ran in an emulated network with a trivial exploit. He also disputes the self-replication framing, noting the models used sharding rather than full propagation, more like a dropper that installs a small coordination tool than a worm copying itself wholesale. A true polymorphic worm would need to retrain its own models, which Anthropic did not attempt.