Anthropic Says Zhipu's Open-Weight GLM-5.3 Nearly Matches Claude at Writing Exploits
The Decoder · rss · 2026-09-30
According to The Decoder, Anthropic's evaluation found that Zhipu's open-weight GLM-5.3 writes cyber exploits nearly as well as Claude Mythos Preview.
The smaller Flash variant assembled a reliable Chrome attack for just $20.40 at Zhipu's API prices. Anthropic also warned the model's safeguards are easy to strip out, and unlocked versions are already circulating.
While Anthropic has its own incentives to sound the alarm, the US agency CAISI's assessment backs up the findings, highlighting the dual-use security risks of open-weight models.
More from Models
- User predicts Anthropic will ship a Claude personal assistant by November — LeviTurk · 2026-09-30
- User hits Claude Opus 5.5 weekly limit fast, juggles AI subscriptions as prices fall — StewartalsopIII · 2026-09-30
- Users report Claude usage limits reset after Dev Day — some got reduced — haltakov · 2026-09-30
- BiliBili Index Team Releases Qwen3.5-Based Translation Models Supporting 150 Languages — AdinaYakup · 2026-09-30
- ChatGPT's rich interactive responses quietly disappeared overnight, user reports — Norddinerezouk · 2026-09-30
- Self-hosting full GLM 5.3 takes an 8-GPU rig, roughly $136k at today's prices — TheZachMueller · 2026-09-30