An 0.7-point edge on review is enough to scare a CISO. Close is not comforting.
CCP-Linked AI Rivals US Anthropic’s Model In Cyber Tests
A CCP-tied Chinese model just matched Anthropic on finding software flaws.
Summary
- Daily Caller reports Zhipu AI (Z.ai), a Chinese firm with CCP ties, claims its GLM-5.3 model beat Anthropic's Mythos 5 at finding software security flaws in a CyberGym-style test.
- GLM-5.3 scored about 84.5 percent on coding review versus 83.8 percent for Mythos 5 in the unverified run cited by the outlet.
- The Chinese model was weaker at turning flaws into working exploits, but the gap on discovery is the headline rivals wanted.
- Zhipu plans an open-source release, which would put the weights in more hands outside US export controls.
- The story lands as Washington still argues how hard to fence advanced Chinese AI.
Commentary
China does not need to win the press release. It needs a good-enough cyber model that ships open and cheap.
Matching Anthropic on flaw-finding is a warning light for every lab and utility that still treats PRC models as toys.
Export controls without enforcement are cosplay. If Party-linked shops can race US labs on security tasks, the fence has holes.
Comments
Open weights from a Party-tied shop are a gift to every crew that cannot buy Claude.
Exploit conversion lagged. Discovery did not. That is still a weapons lab input.
Treat Zhipu like a defense contractor with a Beijing board, not a startup in a hoodie.
Japan's firms already see PRC models undercutting on price. Capability is next.
USAID seminars will not patch a firmware hole found by GLM. Sanctions and compute will.
Daily Caller flagged CCP ties. Demand the bench scores, not a vibes denial from the embassy.
Same rule as ports: if the rival builds the crane, do not rent it for your shipyard.
Ban the weights from federal nets until the ownership chart is clean.
Cyber parity is how blockades get mapped. Do not sleep on it.