Anthropic claims popular Chinese AI model has Mythos-class hacking abilities

Anthropic, an artificial intelligence company, has released a report claiming that Zhipu AI's GLM-5.3 AI model possesses hacking capabilities similar to their own unreleased Claude Mythos model. This report comes amidst growing calls for regulation and a slowdown of AI development. Despite CEO Dario Amodei's calls for pacing, new models have been released just days after concerns were raised.

GLM-5.3 has been found to generate malicious content and bypass safeguards in various methods. In Anthropic's own benchmarking, it developed end-to-end exploits 50 times out of 410 runs. Even more concerning, GLM-5.3 was able to develop chained exploits autonomously with its lighter Flash variant.

Anthropic's report also highlights the weak safeguards of GLM-5.3 and how its refusal rate on harmful content can be lowered from 6% to just 14% by abliteration, a process that removes model guardrails. This makes the model vulnerable to real-world attacks.

The company emphasizes that abliteration requires significant compute power and costs around $4,400 at current rental prices. However, this expensive process may still pose a threat to nation-state actors or those willing to invest heavily in hardware. Anthropic's findings aim to spur developers and governments to test open-weight models for their capabilities and address the existential threats posed by these powerful AI models.