GLM-5.3 was previously identified by Anthropic as nearly equal in cyber-capabilities to its own restricted Claude Mythos Preview, with the two models succeeding in 12% and 14% of exploit tasks respectively. Anthropic found that removing the model's safeguards required 2,200 GPU hours at a cost of roughly $4,400. While Anthropic has called for government safety testing of such capable models, reports suggest the public release of GLM-5.3 has not yet resulted in major real-world cyberattacks.
Tuesday, October 20, 2026
Story
Venice Launches First Anonymous Uncensored Version of GLM-5.3 AI
AI14 days ago
A modified version of Zhipu AI's GLM-5.3 model, stripped of refusal mechanisms, is now accessible to users without authentication through the Venice platform. Known as Abliterated Large V2, the model allows anonymous access to tools for exploit analysis and red-team scenarios. In a separate security demonstration, researcher Prince Chaddha showed that an open 7B model could be backdoored for under $50 to silently steal user credentials with a 100% success rate, highlighting the risk of poisoning the AI supply chain through platforms like Hugging Face.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.