Anthropic says Zhipu's open-weight GLM-5.3 nearly matches Claude Mythos Preview at building exploits
Anthropic's Frontier Red Team evaluated Zhipu AI's open-weight GLM-5.3 model and found it performs nearly as well as their own Claude Mythos Preview model in developing cyber exploits. On Anthropic's internal binary exploitation benchmark, GLM-5.3 achieved full control flow hijacks in 4% of trials, compared to 6% for Claude Mythos Preview. Both models significantly outperformed earlier versions…
Key points
- GLM-5.3 achieved 4% success rate on Anthropic's internal binary exploitation benchmark; Claude Mythos Preview achieved 6%.
- GLM-5.3 built working exploits in 50 of 410 ExploitBench attempts; Mythos Preview managed 56 in 410 attempts (Decoder only).
- Abliterating GLM-5.3-Flash took 2,200 GPU hours at ~$4,400 cost; model requires 306 GB VRAM (Tom's Hardware only).
The reports differ in additional details: only the Decoder article mentions that GLM-5.3 built working exploits in 50 of 410 attempts on the ExploitBench benchmark, while Mythos Preview managed 56 successes in the same number of attempts. The Decoder also notes that Anthropic paired GLM-5.3 with a human expert to discover previously unknown vulnerabilities in a browser's JavaScript engine, chaining them to extract private SSH keys. Tom's Hardware adds that abliterating GLM-5.3-Flash took 2,200 GPU hours at an estimated cost of $4,400, and that the model's weights require 306 GB of VRAM at full-precision FP8.
Model pages: GLM 5.3 → · Claude Opus 5.5 →
The story so far
9 episodes →- Anthropic says Zhipu's open-weight GLM-5.3 nearly matches Claude Mythos Preview at building exploitsthis story
Quoting Anthropic Frontier Red Team
Simon Willison · 29 September 2026
Loading the full article…
This text was published by Simon Willison and written by Simon Willison. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
3sources- Anthropic claims popular Chinese AI model has Mythos-class hacking abilitiesPress · Tom's Hardware ·
- Anthropic says Zhipu's open-weight GLM-5.3 nearly matches Claude Mythos Preview at building exploitsPress · The Decoder ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Anthropic's Claude begins 'dreaming' to reorganize memory in tests · 1 src
- John Langford quits Microsoft, declares AI Frontiers over · 1 src
- Anthropic releases study on Claude usage in scientific research · 3 src
- MIT and BASF release CheMeleon for low-data molecular prediction · 2 src
- Anthropic tests Claude agents in EVE Online trading and sabotage detection · 1 src
Comments
via GitHub Discussions