October 14, 2026
- Anthropic's self-replication study is the concrete version of the "AI worm" question. Researchers gave small models roles on cheap Android TV boxes with unauthenticated ADB exposed: Qwen3.5-9B installed and configured the software to run Llama3.1-8B on eight of the nine hardware types tested, and GLM-5.2 reached all eight nodes in an emulated network in four of five trials. Retaining data from previous attempts raised Qwen3.5-9B's success rate from 45% early to 84% in later generations; using Opus 4.8 as an advisor produced no measurable improvement. Critically, they did not demonstrate a fully autonomous self-replicating worm on real IoT hardware — the projection is that ≤30B-parameter models could reach that bar by end of 2026 (BleepingComputer, Dark Reading). HN commenter evan_a_a points to Anthropic's own write-up as the thing to read rather than the coverage. (discussion) · AI & Model Security
in Three Chained Zero-Days Put Root on a Fully Patched Ivanti VPN