Anthropic claims popular Chinese AI model has Mythos-class hacking abilities (www.tomshardware.com)

Anthropic has released a report asserting that Zhipu AI's GLM-5.3 model possesses dangerous capabilities akin to their own unreleased Claude Mythos AI, particularly in generating malicious content. The findings claim that GLM-5.3 can be misused for cyberattacks due to bypassable safeguards, creating a potential security risk in the AI landscape. In their experiments, GLM-5.3 demonstrated a 4% success rate in performing complex exploits, just slightly below that of Mythos, while it was capable of generating chained exploits at a low cost, making it economically attractive for potential attackers. The significance of this report lies in its timing amidst growing fears about AI safety and calls for regulation. Anthropic is emphasizing the threats posed by open-weight AI models like GLM-5.3 in comparison to their closed-source counterparts, urging developers and governments to evaluate the risks associated with these accessible technologies. While the costs and technical requirements to harness GLM-5.3 for malicious purposes are considerable—potentially requiring extensive computational power—Anthropic's findings spotlight the escalating need for governance and the potential dangers that unchecked AI models can pose in cybersecurity scenarios.