An open-weighted Chinese AI model came surprisingly close to Anthropic's limited model in detecting and exploiting software vulnerabilities.

An open-weighted Chinese AI model came surprisingly close to Anthropic's limited model in detecting and exploiting software vulnerabilities.

The model is GLM-5.3, built by Zhipu AI, or Z.ai.

Anthropic tested it against known vulnerabilities in Chrome's V8 JavaScript engine.

Out of 410 trials, GLM-5.3 produced 50 complete survivors.

Anthropic's Claude Myth preview made it to 56.

Mythos is restricted to vetted cyber security researchers.

GLM-5.3 is open weighted. Anyone can download it.

Anthropic has also found that its protective measures are relatively easy to remove. One technique reportedly beat them 100% of the time, with little loss in ability.

And stripped-down versions of the model appeared online within days.

In controlled testing, Anthropic says GLM-5.3 even discovered previously unknown browser vulnerabilities and patched them together.

Source: SCMP