Anthropic's Frontier Red Team analysis indicates that Zhipu AI's open-weight model, GLM-5.3, demonstrates cyber-exploit capabilities comparable to the Claude Mythos Preview. While Anthropic restricts access to its model through Project Glasswing, GLM-5.3 is publicly available for download.
On the ExploitBench test, GLM-5.3 successfully created working exploits in 50 out of 410 attempts, compared to 56 for Mythos Preview. In internal binary exploitation benchmarks, GLM-5.3 achieved control in 4 percent of tasks, while the Claude model reached 6 percent.
Anthropic researchers demonstrated that GLM-5.3 could identify unknown vulnerabilities in a browser's JavaScript engine and extract private SSH keys. Using the GLM-5.3-Flash variant, the team constructed a functional attack from a Chrome bug in eight hours of model time at a cost of $20.40.
The US agency CAISI identified GLM-5.3 as the most cyber-capable open-weight model, trailing top US models by approximately four months. Anthropic noted that GLM-5.3's security safeguards are easily bypassed, with refusal rates for malicious requests dropping to between 2 and 12 percent after applying abliteration techniques.
The abliteration process required 2,200 GPU hours at a cost of roughly $4,400, though Anthropic estimates experienced teams could perform it for $1,200. The UK's AI Security Institute observed that the capability gap between open and closed models has narrowed to four to seven months.
