Anthropic Warns GLM-5.3 Spreads Advanced Cyber Capabilities, Critics Call It an Ad
量子位 · wechat · 2026-09-30
Anthropic's threat report flags Zhipu's open-weight GLM-5.3 as spreading frontier-level exploitation capabilities with weak guardrails: on ExploitBench it succeeded 50/410 attempts vs 56 for Claude Mythos Preview; researchers built a PAC-bypassing Chrome exploit chain in 8 hours for $20.40 of API calls.
More from Models
- Self-hosting full GLM 5.3 takes an 8-GPU rig, roughly $136k at today's prices — TheZachMueller · 2026-09-30
- Work once delegated to juniors now done by Opus 5.5 in minutes at 95%+ accuracy — TraditionalHome8852 · 2026-09-30
- Qwen3.8-27B, a compact dense model for agents, goes live on Nebius — Alibaba_Qwen · 2026-09-30
- Anthropic Says Zhipu's Open-Weight GLM-5.3 Nearly Matches Claude at Writing Exploits — The Decoder · 2026-09-30
- 5x Pro user review of new model 6.1: stronger than predecessors but much slower on long tasks — Chemical-Agency-3997 · 2026-09-30
- User asks OpenAI's Dot to inspect its own sandbox: 9 CPUs, 9.7GB RAM, no GPU — Caprisuner · 2026-09-30