Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

NewsResearch

Anthropic: China's open-weight GLM-5.3 builds exploits nearly as often as Mythos Preview, and its refusals come off for about $4,400

Anthropic's Frontier Red Team tested GLM-5.3, an open-weight model from China's Zhipu AI (Z.ai), and reported on September 29 that it built working end-to-end exploits in 50 of 410 attempts on ExploitBench, against 56 for Claude Mythos Preview, and achieved full control-flow hijacks in 4 percent of trials on an internal benchmark, against 6 percent; earlier models such as Claude Opus 4.6 and GLM-5.2 succeeded in none. In human-guided sessions GLM-5.3 found several unknown flaws in a browser's JavaScript engine and chained them into a web page that reads arbitrary files from the visitor's computer, and the smaller GLM-5.3-Flash built an exploit for a known Chrome flaw (CVE-2026-11645) with 20 minutes of human attention and eight hours of work, for $20.40. Its safeguards refused bare malicious requests, but a deceptive cover story got it to engage 64 percent of the time and prefilled reasoning 92 percent; abliteration, editing the public weights to remove refusals, took Anthropic's team about 2,200 GPU hours (roughly $4,400) and cut refusals from above 90 percent to about 2 to 3 percent on two benchmarks and 12 percent on a third, without significantly reducing capability. Anthropic says none of these techniques got its safeguarded Claude models to carry out the harmful tasks it tested, and that GLM-5.3 "will likely give malicious actors access to capabilities that will allow them to find and exploit cyber vulnerabilities without meaningful restrictions". On September 17 NIST's Center for AI Standards and Innovation had called GLM-5.3 "the most cyber-capable open-weight model released to date", while finding it significantly behind US frontier models, by about four months. The safeguard tests ran in a simulation in which no model-generated code was executed.

Why it matters

Exploit-building that recently needed a restricted frontier model is now in weights anyone can download and strip of refusals for the price of a laptop. Plan for attackers who have it: shorten patch times for internet-facing systems, and treat a published proof of concept as the start of exploitation.

Sources

Read it with the rest of the news in the Learning Center