Open Weight GLM-5.3 Almost Equals Claude In Writing Exploits
GLM-5.3 (an open-weight model) is almost as good as Claude Mythos Preview for exploit generation. Someone even made a functional Chrome exploit for $20.
GLM-5.3, an open weight model developed by Zhipu, is, according to Anthropic, “very close” to Claude Mythos Preview in terms of exploit generation ability (exploits are pieces of code that hack into other software and gain control). In a test on ExploitBench, GLM-5.3 generated a functional exploit 50 times out of 410 attempts, compared to 56 times for Mythos Preview. When testing on a binary vulnerability task (a task involving identifying vulnerabilities in compiled code), GLM-5.3 completely hijacked the target program 4% of the time, while Mythos Preview did so 6% of the time. According to the CAISI agency, GLM-5.3 is currently the most threatening open weight model in terms of cybersecurity, lagging approximately four months behind the top U.S. models.
The disadvantage of open weight models is that they can be easily deprived of almost all their safety constraints within a few days due to their openness, allowing anyone to customize them for personal use. It cost between $1,200 and $4,400 to “uncensor” GLM-5.3, reducing its rejection rate of malicious requests from 90% to 2–12%. Generating a functional Chrome exploit via Zhipu’s API costs $20.40. This is why I prefer to use closed weight models like Claude in my daily life — because I know the company will continue to bear responsibility for the model even after its launch. However, once the safety budget has been exhausted, there is no longer any responsibility for an open weight model.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles