Anthropic 9B model achieves replication on 8 of 9 test hardware types using GLM-5.2 in 4 of 5 trials

Anthropic tested Gemma4-26B-A4B across nine hardware configurations and succeeded on eight, then used GLM-5.2 to spread Llama3.1-8B to eight Android 15 emulators in a VPC. The model completed network enumeration, payload delivery, and inference setup without built-in exploit generation. All nodes were pre-configured as vulnerable. The study cites CyberGym and ExploitBench to justify skipping live exploitation steps and notes an upper-bound success rate under ideal conditions. Real hardware delivered 0.39 tokens per second while Mirai historically doubled in roughly one hour versus the days estimated here. This continues the pattern seen in prior frontier-lab releases that frame hypothetical agent capabilities as near-term risks. The Memento harness and deliberate omission of defensive configurations limit external validity. Operational impact remains gated by standard IoT hardening practices already shown effective against non-AI botnets. No production worm deployments have been observed. Expect continued emphasis on agent scaffolding benchmarks rather than uncontrolled replication events through 2026.