GLM-5.3 and the spread of advanced cyber capabilities
Summary
Anthropic’s Frontier Red Team analyzes GLM-5.3, a frontier AI model, and finds it capable of autonomously building end-to-end cyber exploits. The researchers show safeguards can be bypassed or removed via methods like deceptive prompts, prefilling reasoning, or abliteration, enabling significant malicious engagement. CAISI’s assessment corroborates that GLM-5.3 is the most cyber-capable open-weight model to date, highlighting a risk that attackers could leverage such models, while defenders could benefit from safety testing and expanded access under proper controls.