GLM-5.3 and the spread of advanced cyber capabilities
Anthropic finds Zhipu's open-weight GLM-5.3 autonomously builds end-to-end exploits and its safeguards can be bypassed 64-100% of the time.
Anthropic analysis finds Zhipu AI's GLM-5.3, released without meaningful safeguards, autonomously develops end-to-end exploits: 50 of 410 attempts on ExploitBench (vs. 56 of 410 for Claude Mythos Preview) and full control-flow hijacks on 4% of OSS-Fuzz binary exploitation tasks, where earlier models like GLM-5.2 and Claude Opus 4.6 scored zero. Simple techniques bypassed GLM-5.3's safeguards 64-100% of the time in simulated tests. NIST's CAISI separately assessed GLM-5.3 as the most cyber-capable open-weight model to date, lagging the U.S. frontier by about four months. During researcher-driven testing, GLM-5.3 chained multiple discovered zero-days in a popular web browser to steal an SSH private key.
- GLM-5.3 matches Claude Mythos Preview on end-to-end exploit development benchmarks
- Safeguards bypassed 64-100% of the time with simple techniques; anyone can download the weights
- NIST CAISI rates it most cyber-capable open-weight model, ~4 months behind US frontier
- Researchers used it to discover chained zero-days in a popular web browser in under a day
- Tags
- ai-safetyglm-5.3zhipu-aicyber-capabilitiessafeguard-bypassexploit-generationopen-weightsanthropic
- AI models
- GLM-5.3Claude Mythos Preview
From the original article
10 min read at anthropic.com
Five months ago, we announced Claude Mythos Preview, the first AI model that could autonomously build sophisticated, end-to-end cyber exploits. The rapid rate of improvement in AI suggested to us that this ability would eventually proliferate to many other models, making it much easier for malicious cyber actors to launch highly impactful cyberattacks.
By Philpax · anthropic.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.