A Wall Street Journal report highlighted Zhipu AI's GLM-5.2 matching top U.S. models like Anthropic's Mythos at spotting security bugs in targeted Semgrep tests, yet Miles Brundage counters that this falls short of proving equivalence and stakes a 100-to-1 wager on upcoming UK AISI cyber range results where GLM-5.2 has no public scores.
The #1 article on the Wall Street Journal claims that GLM-5.2 matches Mythos at finding security bugs.
This is almost certainly completely incorrect.
I am willing to put up $100 for $1 that GLM-5.2 will score below Mythos and GPT-5.5 at UK AISI's cyber range.
WSJ has been...
Semgrep evaluations focus on isolated vulnerability detection tasks while AISI ranges require autonomous multi-step network operations, leaving open whether current GLM-5.2 numbers translate to the longer-horizon scenarios that matter for Brundage's bet.
Without public UK AISI results on GLM-5.2, the dispute rests on whether its Semgrep performance will hold up against Mythos and GPT-5.5 when both models face the same end-to-end simulated enterprise breaches.
Positive users defended WSJ's economic reporting or welcomed scrutiny of AI hype around open models, while negative users criticized the GLM-5.2 bug detection coverage as inaccurate and a dumpster fire that worsens public misunderstanding.
No Digg Deeper questions have been answered for this story yet.
here is the article, the only one i found: https://www.wsj.com/tech/ai/chinese-ai-anthropic-mythos-cybersecurity-574b02c2
Supposedly, "a new model from" from zAI is said to be at least as strong as Fable5 in cybersecurity-related aspects.
I did some research and only came across a Wall Street Journal article, which, however, does not refer to a new model, but to GLM 5.2 as a relatively new model that was released recently. So either GLM 5.2 is stronger than people think, or the news being circulated is misleading.
According to WSJ, Zhipu AI’s GLM-5.2 can match top US models in some bug-finding scenarios, and China’s 360 Security says its new Tulongfeng tool is comparable to Anthropic’s Mythos.
...doing a really bad job in general when it comes to reporting on AI model's capabilities.
Please @bobmcmillan @AmrithRamkumar you should take down this article. I'm sorry. The headline takeaway is just wrong. If you talk to researchers at the UK @AISecurityInst, I'm sure...
@Tim_Hua_ same tbh. just makes me even more concerned than i was before about lack of public ai understanding
@bobmcmillan @AmrithRamkumar @AISecurityInst ...they're agree with my assessment here.
@space_colonist Nah I think their Econ and political reporting are pretty good.
Maybe to operationalize this a bit more, “better” is measured either at the point where the model plateaus, or anywhere in like the high compute (>$50 dollar per task) region. If GLM 5.2 does better than mythos when you spend like $5, that doesn’t count.
@Tim_Hua_ welcome to the Gell-Mann amnesia effect
I’ll also do 200:1 for Mythos proper (not preview.)
@Tim_Hua_ finally we are seeing through the hype lately with open source models
@Tim_Hua_ like i thought people would at least sort of keep up but it’s a total dumpster fire
@Tim_Hua_ interesting bet; data will tell the real story
