Anthropic's team removed the model's refusal mechanisms for approximately $4,400, spending 2,200 GPU hours on a process called abliteration. This reduction in guardrails dropped the refusal rate from over 90% to as low as 2% without significantly impacting general intelligence or scientific capabilities. Earlier models, including GLM-5.2 and Claude Opus 4.6, failed to create any successful exploits in these trials, though GLM-5.3 managed full control flow hijacks in 4% of cases. OpenAI added GLM-5.3 to its Codex subscriptions on the same day the safety research was published.
Tuesday, October 20, 2026
Story
Open GLM-5.3 Rivals Restricted AI in Autonomous Cyber-Exploits
AI21 days ago
Zhipu AI's latest open-weight model can independently create functional hacking tools and navigate complex security sandboxes, according to research from Anthropic. The GLM-5.3 model built working browser exploits in 50 of 410 trials, nearly matching the 56 successes of Anthropic's own restricted Claude Mythos Preview. During sandboxed testing, GLM-5.3 discovered unknown browser vulnerabilities within a single day and chained them into a webpage to steal an SSH private key. Anthropic warns that unlike proprietary frontier models, this downloadable version was released without robust safeguards, which analysts bypassed between 64% and 100% of the time in simulated tests.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.