Z.ai says GLM-5.3 scores 84.5% on CyberGym, vs. Mythos 5's 83.8%, and its most sensitive cybersecurity functions will only be available to verified users

Chinese AI startup Z.ai said on Friday its open-source GLM-5.3 model had neared Anthropic's restricted Mythos 5 in identifying software …

Reuters 2026-08-15

Context & Ripple Effects

Z.ai has steadily positioned its GLM line as open-weight competition in reasoning, coding and agentic work, beginning with its GLM-5 flagship launch and extending that claim with an MIT-licensed GLM-5.1 release. GLM-5.3 moves that performance narrative into cybersecurity.

The reported CyberGym comparison matters because Z.ai is pairing a near-peer benchmark claim against Anthropic's restricted Mythos 5 with verified-user controls for its most sensitive functions. That makes access policy part of the product distinction, not merely a post-release safeguard.

First-order effects

Z.ai can market GLM-5.3's 84.5% CyberGym result against Mythos 5's 83.8% while limiting its highest-risk cybersecurity capabilities to verified users.
Users seeking the sensitive functions must pass Z.ai's verification gate, separating access to those capabilities from the broader availability implied by the GLM line's open-source positioning.

Second-order effects

Anthropic's restricted Mythos 5 and Z.ai's verification policy make cybersecurity-model comparisons increasingly about both measured capability and the terms under which that capability is reachable.
Developers and enterprise buyers evaluating GLM-5.3 must account for identity verification as part of deployment planning, rather than treating model access as uniform across cybersecurity tasks.

Third-order effects

If leading labs continue pairing cyber benchmarks with differentiated access tiers, cybersecurity capability is likely to become a governed product layer even where underlying model distribution is comparatively open.
The competitive boundary may shift from whether a lab releases a model to how it authenticates, monitors and limits access to its most sensitive functions.

The trend: Frontier AI labs are turning high-risk cybersecurity capability into a governed-access layer while continuing to compete publicly on benchmark performance.

Discussion

@joshua_saxe Joshua Saxe on x
Really cool to see a http://z.ai / GLM engineer demonstrating defensive cyber capabilities of their new model. Feels like there's a real possibility that if the cyber guardrails on the American closed models don't change the security community will move over to these models at s…
@zaddyzaddy @zaddyzaddy on x
The cybersecurity doom-mongering from the big labs is about to lose its force. GLM-5.3 has shown that existing models can become significantly more capable on cyber tasks with additional RL training. So what stops others from building on GLM-5.2 and pushing its cybersecurity ca…
@hesamation @hesamation on x
While Anthropic and OpenAI put cyber defense behind safeguards, GLM 5.3 is specifically trained for coding and cyber defense. Wow.
@samhogan Sam Hogan on x
GLM 5.3 benchmarks look great, pretty substantial improvement over 5.2 on coding and cyber
@elshayib_ @elshayib_ on x
GLM-5.3 Sucks come on @Zai_org you're doing it wrong, you can't say your model is good for cyber defense with 0 Sandbox escapes, come on step up your game.
@zixuanli_ Zixuan Li on x
Building on GLM-5.2's contributions to cyber defense, GLM-5.3 delivers substantially stronger capabilities in vulnerability discovery, exploit analysis, and complex, multistep security tasks in realistic environments. These advances can help defenders identify weaknesses earlier,
@_nathancalvin Nathan Calvin on x
“API access and open weights will be released in stages following rigorous safety evaluations.” GLM 5.3 seems to be extremely good at cyber defense and cyber offense, curious what rigorous safety evaluations entails
@druce.ai @druce.ai on bluesky
China's Z.ai says new model nears Anthropic's Mythos 5 in cyberdefense tests, security abilities will be restricted to a “trusted access” program

Chronicles