Technology
Anthropic flags cyber risks in China’s GLM-5.3
Business Standard··1 Oct
Anthropic said China’s open-weight GLM-5.3 from Z.ai crossed a threshold in AI-assisted cyber capability testing. On ExploitBench covering Chrome’s V8 engine, GLM-5.3 built end-to-end exploits in 50 of 410 attempts. Claude Mythos Preview scored 56 of 410 on the same set. Anthropic said open weights make safety controls easier to weaken than in closed Claude access.
Prism
What It Means For You
- Open-weight release means users can download and alter model behaviour outside a vendor’s hosted controls.
- Anthropic said simple prompt framing raised engagement on harmful cyber requests in its tests.
- The company ran the evaluations in isolated sandboxes rather than on live systems.
What's Happening
- Zhipu AI, known outside China as Z.ai, released GLM-5.3 in August as a frontier model.
- On Anthropic’s Binary Exploitation benchmark, GLM-5.3 achieved full control-flow hijacks in 4 percent of 100 tasks.
- A GLM-5.3-Flash trial chaining a disclosed Chrome flaw cost $20.40 at Zhipu API prices, Anthropic said.
How the Models Were Compared
- Anthropic compared GLM-5.3 with Claude Mythos Preview and Claude Opus Preview on exploit-development tasks.
- Researchers said an abliterated GLM-5.3 cut refusal rates sharply while keeping much general capability.
- Claude models remain available through controlled access without releasing underlying weights to users.
all-news
