Anthropic Reports Zhipu's GLM-5.3 Model Competes with Claude in Exploit Generation

2026-10-06

Anthropic states that Zhipu's open-weight GLM-5.3 model demonstrates capabilities in generating cyber exploits comparable to its Claude Mythos Preview. A smaller variant, GLM-5.3 Flash, reportedly created a Chrome attack for approximately $20.40 using Zhipu's API.

VERA Brief

AI-generated. Grounded in the article and its cited sources.

Anthropic reports that Zhipu AI's GLM-5.3 model shows capabilities in generating cyber exploits similar to Anthropic's Claude Mythos Preview. A variant of this model reportedly created a Chrome attack for a low cost, raising concerns about AI safety training and potential misuse.

Key facts

Anthropic has reported that Zhipu AI's GLM-5.3 model is comparable to its Claude Mythos Preview in generating cyber exploits.
A smaller variant, GLM-5.3 Flash, reportedly created a Chrome attack for approximately $20.40 using Zhipu's API.
The model's built-in safeguards are reportedly susceptible to removal, with unprotected versions already available.
Anthropic's assessment is reportedly corroborated by the U.S. agency CAISI.

Source: The Decoder

Reported by VERA Newswire.