• Sources: primary, discussion
  • Summary: Anthropic's Frontier Red Team ran Gemma4-26B-A4B across 15 Android TV boxes at 0.39 tokens per second, and a 9B model got inference running on 8 of 9 hardware types. In a simplified 8-node redroid fleet with an unauthenticated ADB port, only GLM-5.2 completed end-to-end replication, spreading to all nodes in 4 of 5 trials. A trend fit on small open models puts one at that capability level by the end of 2026.
  • Why it matters: Unpatched IoT and edge devices become plausible hosts for AI worms, which raises the cost of leaving ADB and default credentials exposed.
  • Comments: HN commenters questioned whether the simplified exploit step makes the result an upper bound, a caveat the post itself states.