IC Technology & Cyber — 2026-07-16

CISA Official Says Chinese AI Model GLM 5.2 Cybersecurity Capabilities Surpass Expectations and Rival Anthropic Mythos

BLUFCISA's own benchmarking of Chinese open-weight models as peer-competitive on offensive cyber tasks erodes the strategic premise underlying US AI export controls before those controls have taken full effect.

Jay Gazlay, deputy associate director of CISA's vulnerability management section, said at a July 15 CSIS event that GLM 5.2, an open-source frontier AI model developed by China-based Z.ai, showed "amazingly effective" cybersecurity capabilities that he found "stunning" 1. Gazlay said his testing of GLM 5.2 "even took my breath away" compared with Claude Mythos, the Anthropic model CISA gained access to in June 1. Gazlay separately said CISA tested an unnamed frontier AI model's ability to convert a recently disclosed open-source vulnerability into a durable remote-code-execution exploit, without naming the model 1. Former Pentagon cyber official Mieke Eoyang, speaking at the same event, said the emergence of such models "change[s] our calculus" on risks from Chinese technology adoption, citing Beijing's vulnerability-disclosure policies and the Commerce Department's June 30 lifting of export controls on two Anthropic models 1.

Analysis
CISA's on-record comparison, treating GLM 5.2 as competitive with top-tier US commercial systems on offensive cyber tasks, undercuts claims that export controls preserve a meaningful capability gap. CISA's own use of a frontier model to convert a routine open-source flaw into a durable remote-code-execution exploit shows the agency deploying AI-assisted vulnerability research operationally, not just tracking adversary use, though that model's unnamed status leaves attribution and export-control implications unresolved. Eoyang's linkage of this convergence to Commerce's June 30 reversal on Anthropic's export controls reframes the debate around offensive-cyber risk rather than intellectual property. The comparison rests on a single account of one CSIS panel, with informal testing rather than published benchmarks behind it, and also doubles as an argument for expanded interagency authority over frontier AI models.

2 sources

CISA official Gazlay sees opportunity to address China-developed frontier AI models through interagency efforts - Inside Cybersecurity
The Future of AI-Driven Cyber Defense - CSIS