Anthropic's red team says small open AI models may be months away from copying themselves across cheap smart devices, with no data centre needed

Anthropic's Frontier Red Team published a research post on 12 October, written by William Zhou and Daniel Freeman, on whether open-weight models of 30 billion parameters or fewer could set up their own inference on cheap hardware and spread from device to device. It says: “Our experiments suggest that small AI models may be months away from being capable enough to self-replicate” on everyday internet-connected devices such as smart ovens and off-brand Android TV boxes.

In a simulated worm test on a firewalled network of eight cloud nodes running Android emulators, Z.ai's open model GLM-5.2 spread to all eight nodes in four of five trials and to seven in the fifth; the other models tested failed. A 9-billion-parameter model got inference running on eight of nine hardware types. The authors say the hacking step was made easy (an unauthenticated debug port), so the success rates are upper bounds. On a general capability index, they estimate small models could reach GLM-5.2's level by the end of this year, and call that a rough estimate.

The post says: “We don’t yet have a full rogue replicator”. Because such a replicator would have no central off switch, it says internet providers' traffic monitoring could spot the multi-gigabyte model transfers and quarantine infected customers, and it discusses decoy devices to catch a replicator and more secure defaults from manufacturers.

The post drew about 130 outlets, among them the FT, CNBC and Al Jazeera, the most-covered AI story of the week that the Switchboard had not run as a story.

• Claimed Anthropic says small AI models “may be months away from being capable enough to self-replicate”. Anthropic
• Confirmed In Anthropic's simplified test, GLM-5.2 spread to 8 of 8 nodes in 4 of 5 trials and 7 of 8 in the fifth. Anthropic
• Claimed Anthropic says it does not yet have a full rogue replicator and that the exploit step was simplified, so rates are upper bounds. Anthropic