Anthropic flags Chinese GLM-5.3 model for elite hacking capabilities

By Splitfeed EditorialLast updated Oct 1, 2026, 06:00 UTC

On September 29, 2026, Anthropic published a report titled "GLM-5.3 and the spread of advanced cyber capabilities," asserting that GLM-5.3, a Chinese AI model it describes as popular and widely available, has crossed into what Anthropic calls Mythos-class hacking ability. That classification places GLM-5.3 among the most capable systems Anthropic has evaluated for offensive cybersecurity tasks. The report does not appear to be a response to an incident but rather a proactive assessment, and no response from GLM-5.3's developers has appeared in the available sources.

The significance of the report lies in what the Mythos-class designation implies. Anthropic has developed its own internal taxonomy for rating AI cyber risk, and placing a widely distributed model at the elite tier of that scale is a claim that, if accepted, would mean serious offensive hacking capabilities are no longer confined to closed or tightly controlled systems. The framing of the report around "the spread of advanced cyber capabilities" suggests Anthropic is making a broader argument about proliferation, not just about one model in isolation. What remains genuinely in dispute is whether Anthropic's evaluation methodology is sound, whether the Mythos-class label maps onto real-world threat levels in a meaningful way, and whether a commercial AI safety company is the appropriate or unbiased party to make such determinations about a competitor nation's technology.

None of the three camps have published reactions to this story yet. The Pro-AI camp would typically push back on alarmist framing, questioning whether the benchmark reflects actual deployable risk or overstates what any model can do in practice. The Anti-AI camp would typically treat the findings as confirmation that frontier-level dangers are spreading faster than governance can respond, and would likely call for stricter controls on model access. The Middle Ground camp would typically call for independent verification of Anthropic's methodology and caution against reading geopolitical motivations into either the report or any dismissals of it.

The immediate questions to watch are whether GLM-5.3's developers issue a formal response, whether independent security researchers attempt to replicate Anthropic's evaluation, and whether the report prompts any regulatory or policy response in the United States or elsewhere regarding access to models assessed at this capability tier.