Anthropic red team: small models may be able to self-replicate on smart devices within months

發生什麼事

In an 8-node simulated network, Anthropic's research had GLM-5.2 complete the full replication process through unauthenticated ADB ports, infecting every node in 4 of 5 runs. The researchers also chained 15 TV boxes costing about $55 each to run a model at 0.39 tokens per second, and a 9B model got inference running on its own on 8 of 9 hardware types (Anthropic's article). The article also cites past experience with the Mirai botnet: when network providers quarantined and notified infected users, 92% of infections were cleared within 14 days, versus only 74% without notification, and Anthropic thinks the same approach should work against AI worms. An AI-safety advocacy account relayed the findings in a more alarmed tone than the original (@AISafetyMemes); an account focused on checking primary sources pointed out that the experiment simplified the intrusion steps and that the end-of-year timeline is only a projection (@EarlySignalX).