THE TEST WORM HOPPED DEVICE TO DEVICE AND COPIED ITSELF — 8 OF 8 NODES, 4 OF 5 TRIALS

In Anthropic's self-replication tests, the open-weight model GLM-5.2 was the only model to complete the full chain end to end, spreading to all eight test nodes in four out of five trials. The test broke replication into scanning for vulnerable devices, gaining access, setting up the AI environment on unfamiliar hardware and spreading further. The exploit step was deliberately kept trivial, and Anthropic describes the result as an upper bound rather than a full rogue replicator in the wild.

Other models fell short of chaining every step, though a 9-billion-parameter model set up an environment running Llama3.1-8B on eight of nine device types, and copies that passed notes about successful and failed strategies to successors improved markedly over generations. Anthropic also showed a 26-billion-parameter model running across 15 cheap Android TV boxes. No AI worm has been found in the wild, and the warning is that small models could reach this capability by the end of 2026.