Anthropic: GLM-5.3 safeguards bypassed 64-100% of the time in cyber tests

StefanoGogioso · x · 2026-09-30

Anthropic published an assessment of Zhipu AI's GLM-5.3, finding it capable of autonomously building end-to-end cyber exploits like Claude Mythos Preview — but with weak safeguards. In simulated tests, attackers bypassed GLM-5.3's protections 64-100% of the time with simple techniques, while the same attacks failed against safeguarded Claude models. NIST's CAISI separately called GLM-5.3 "the most cyber-capable open-weight model released to date" on Sept 17. Anthropic assesses the lax safeguards significantly raise malicious actors' cyber capabilities, though defenders could also benefit.

Related event: Anthropic Report: Open-Source GLM-5.3 Near-Frontier in Cyber Attack Capability, Sparking Debate(38 posts)→

Original post →

More from Models

Models channel →