GLM-5.3: Open-Weight Model Approaches Frontier Cyber Capabilities, Anthropic Research Shows
GLM-5.3: Open-Weight Model Approaches Frontier Cyber Capabilities, Anthropic Research Shows
Anthropic’s cybersecurity research says Chinese open-weight model GLM-5.3 can create browser exploits at near-frontier levels and bypass guardrails with simple techniques.
The Chinese model GLM-5.3 is already close to Mythos-level capabilities. That assessment comes from Anthropic itself—at least in cybersecurity.
In its security research, Anthropic writes that GLM-5.3 poses a threat because of its high capabilities and because the model is available to everyone.
ExploitBench: GLM-5.3 produced working browser exploits in 50 out of 410 tasks, while Mythos Preview did so in 56 out of 410. The previous generation of open models scored almost zero. Guardrails are bypassed with fairly simple methods: posing as a red team — 64%, prefilling reasoning — 92%, removing the refusal direction — 100%. With Claude API guardrails enabled — 0%. In a controlled experiment, GLM-5.3 also discovered unknown browser vulnerabilities, assembled them into an exploit chain, and created a web page capable of reading files from a test machine. NIST’s conclusion: GLM-5.3 is the strongest open-weight model in cybersecurity capabilities today. On average, it lags behind the US frontier by about four months.
It is also notable that when an OpenAI model compromised Hugging Face in the summer, it was GLM-5.2 that helped protect open infrastructure. Anthropic and OpenAI declined to help, citing security.
They gave those models an excellent free advertising campaign.
It is unclear whether the companies realize it, but the effect may end up being the opposite of what they expect.
Anthropic research on GLM-5.3 and the spread of advanced cyber capabilities
Why it matters
AnalysisThis matters because open-weight AI models can rapidly spread advanced offensive capabilities beyond controlled frontier labs, while simple guardrail bypasses show that software-only safety limits may be insufficient.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Anthropic Opus 5.5 safeguards bypassed via Opus 4.6 subagent
Opus 5.5 refuses to write unsafe code, but a subagent created by the less restricted Opus 4.6 can significantly weaken its safeguards.
Claude Haiku (Fast Compact Models by Anthropic)
The most compact and fastest model in the Anthropic lineup, designed for lightning-fast responses, mass text classification, data extraction, and agent routing at minimal cost.
Claude Opus (Flagship AI Model by Anthropic)
The most powerful heavyweight model in the Anthropic family, designed for complex philosophical and scientific analysis, multi-layered text generation, and deep contextual understanding.
Anthropic Publishes Guide to Building Always-On AI Agents
Anthropic has released a new guide covering the core mechanisms behind continuously active AI agents, including event handling, timers, memory, and task continuity.