Zvi debunks WSJ's false claim that China's GLM-5.2 matches Anthropic's Mythos

The Wall Street Journal published a headline stating 'China Has Matched Anthropic in Cybersecurity, Resetting AI Race,' which Zvi calls 'obvious nonsense.' The article claims Zhipu AI's GLM-5.2 matches Mythos on certain security bug-finding tasks. Zvi points out that this is only true for the 'easy ones'—where the model is pointed at the correct code subsection and given extra resources. He emphasizes that Mythos's real differentiator is its ability to autonomously identify vulnerabilities across a codebase, chain multiple seemingly unrelated bugs into full exploits, and do so at scale. Neither GPT-5.6 Sol, Opus 4.8, nor GLM-5.2 can replicate that.

Zvi also criticizes the WSJ for misleading reporting that fuels unnecessary regulatory panic. He notes that the 'matching' claim only holds when benchmarks are artificially narrowed, ignoring the autonomous exploit generation that makes Mythos uniquely dangerous. The article warns against Gell-Mann Amnesia—if a major outlet can get this crucial AI safety story so wrong, what else are they misreporting? The piece concludes that while open models like GLM-5.2 are impressive, they are not a credible threat to Anthropic's lead in cybersecurity AI, and releasing Mythos publicly would be a serious mistake.