Home›Global Markets›China›Anthropic warns Z.ai GLM-5.3 has strong hacking abilit…

Anthropic warns Z.ai GLM-5.3 has strong hacking ability and weak safeguards

In Anthropic’s test of end-to-end cyber exploits, GLM-5.3 succeeded in 50 of 410 attempts.

Reported · Written up by our desk · China

Anthropic has warned that China’s Z.ai poses a potential cybersecurity risk through its open-weight model GLM-5.3, citing a combination of strong hacking capabilities and weak safety constraints.

According to a report released Tuesday, Anthropic tested GLM-5.3’s ability to build end-to-end cyber exploits and said the model completed 50 of 410 exploit attempts, compared with 56 for Claude Mythos Preview, a frontier model restricted to vetted users.

Anthropic said GLM-5.3 nearly matches Anthropic’s most advanced model in cyber capability, but falls short on safeguards, raising the risk the system could be co-opted by bad actors.

More like this

Sources