Anthropic: Z.ai's GLM-5.3 Nears Its Frontier on Cyberattacks: what are the key facts?

Anthropic published a report testing the cyberattack capabilities of GLM-5.3, an open-weight model from the Chinese company Z.ai.

Anthropic: Z.ai's GLM-5.3 Nears Its Frontier on Cyberattacks: what are the key facts?

GLM-5.3 succeeded in 50 of 410 exploit attempts, while Claude Mythos Preview succeeded in 56.

Anthropic: Z.ai's GLM-5.3 Nears Its Frontier on Cyberattacks: what are the key facts?

Anthropic says GLM-5.3's cyber capabilities approach those of its most advanced model, but its safety guardrails are far weaker.

Anthropic: Z.ai's GLM-5.3 Nears Its Frontier on Cyberattacks: what are the key facts?

Anthropic warns the model could be exploited by malicious actors, posing a greater risk.

Anthropic: Z.ai's GLM-5.3 Nears Its Frontier on Cyberattacks: what are the key facts?

Claude Mythos Preview is restricted to vetted users, whereas GLM-5.3 is an open-weight model.

Physix Frontier · News Briefing Card (enbrief · Oct 1, 2026)

Anthropic: Z.ai's GLM-5.3 Nears Its Frontier on Cyberattacks

KEY FACTS

Anthropic published a report testing the cyberattack capabilities of GLM-5.3, an open-weight model from the Chinese company Z.ai.
GLM-5.3 succeeded in 50 of 410 exploit attempts, while Claude Mythos Preview succeeded in 56.
Anthropic says GLM-5.3's cyber capabilities approach those of its most advanced model, but its safety guardrails are far weaker.
Anthropic warns the model could be exploited by malicious actors, posing a greater risk.
Claude Mythos Preview is restricted to vetted users, whereas GLM-5.3 is an open-weight model.

KEY DATA

50/410GLM-5.3 exploit successes

56/410Claude Mythos Previe

PHYSIX OBSERVATION

The capabilities of open-weight models are rapidly closing in on closed-source frontiers, yet the safety guardrail gap is stark. Anthropic's move is both a technical warning and a contest for competitive narrative power. For the industry, the disconnect between model capability and safety governance will become a regulatory focal point, and the open-source community needs to prove its controllability.

Source: enbrief original report ↗