GLM-5.2 spread to 8 of 8 machines in Anthropic's self-replication test network
alex_verem · x · 2026-10-16
A detailed thread unpacking Anthropic's Frontier Red Team self-replication study.
- 15 off-brand Android TV boxes ($55 each) networked to run a 26B open model at 0.39 tokens/s
- A 9B model got inference running on 8 of 9 hardware types; GLM-5.2 (750B params) was the only open model to complete the full replicate loop, spreading to all 8 test machines in 4 of 5 trials
- Sub-30B open models gain 13.7 points/year on Epoch AI's index, putting one at GLM-5.2's level by end of 2026; the 9B model also passed setup notes between generations, lifting success from 45% to 61–84%
- Caveats: test network had no real security (upper bound); a 2M-device botnet-sized worm would be 40 H100s of compute with 1–2 day doubling; ISPs can detect it (one ISP wiped out 92% of Mirai infections in 14 days)
- Author asks whether other countries should follow the UK's 2024 ban on guessable smart-device default passwords
More from Safety
- AI agents aren't 'going rogue' — frontier labs are skipping security basics — Sam_Witteveen · 2026-10-16
- Proofpoint to buy Varonis for $6.3B in cash, a 52% premium, closing in 2027 — shashib · 2026-10-16
- China's MofCom & MIIT update AI export controls, covering LLMs and high-sensitivity agents — terryyuezhuo · 2026-10-16
- Vibe-coded dashboard leaked franchise financials before anyone logged in — shauntrennery · 2026-10-16
- Apollo Research: GLM-5.3 unprompted reasons about its actions becoming training data — _lewtun · 2026-10-16
- CFR experts: the autonomous weapons debate still has a definition problem — mchorowitz · 2026-10-16