Sam's News — anthropic — 2026-10-01

Anthropic

8 GLM-5.3 Released With Weak Safeguards Against Cyberattack Capabilities

Anthropic's analysis reveals that Zhipu AI's GLM-5.3 model has strong autonomous cyber-exploit capabilities but lax safeguards, with attackers able to bypass security measures 64–100% of the time. Unlike Claude models released with robust protections, GLM-5.3 is freely downloadable, significantly increasing malicious actors' access to advanced cyber capabilities.

• GLM-5.3 can develop end-to-end exploits autonomously, matching the leap between Claude Opus 4.6 and Claude Mythos Preview
• Attackers bypass GLM-5.3 safeguards 64–100% of the time using simple techniques, compared to unsuccessful attacks against Claude models
• NIST's Center for AI Standards and Innovation assessed GLM-5.3 as 'the most cyber-capable open-weight model released to date,' lagging US frontier by ~4 months
• GLM-5.3 is publicly downloadable without restrictions, unlike US frontier models limited to vetted users
• Anthropic had previously released Claude Mythos Preview in limited form through Project Glasswing, enabling defenders to find 10,000+ vulnerabilities before malicious access

Sources: Anthropic AI Web Searched