Hacks & Exploits
Anthropic Raises Alarm Over Cybersecurity Risks of China's GLM-5.3 AI Model
On September 29, 2026, Anthropic's Frontier Red Team released a report detailing the cybersecurity risks posed by GLM-5.3, an open-weight AI model from the Chinese lab Z.ai. The report indicates that the model could be exploited by anyone who downloads and modifies it, raising concerns about its potential use in cyberattacks.
In testing, GLM-5.3 successfully generated working exploits in approximately 12% of attempts, a figure that closely matches Anthropic's own Claude Mythos Preview. The report highlighted several methods used to bypass GLM-5.3's safety protocols, including deceptive prompts and a technique referred to as 'abliteration,' which allowed for a 100% engagement rate by removing the model's refusal capabilities.
Anthropic's findings suggest that the cost of stripping safety features from GLM-5.3 is relatively low, at around $4,400, while still retaining most of the model's functionality. This raises questions about the effectiveness of built-in safety measures for open-weight AI systems.
Furthermore, the National Institute of Standards and Technology (NIST) rated GLM-5.3 as the most capable open-weight model to date, although it still lags behind U.S. frontier models in terms of cyber capabilities.
Anthropic CEO Dario Amodei has expressed concerns about the threats posed by models that can evade safety protocols but has not advocated for outright bans. Instead, he supports targeted safety testing for advanced AI models, regardless of their open or closed status.
FAQ
What is GLM-5.3?
GLM-5.3 is an open-weight AI model developed by the Chinese lab Z.ai, which has raised cybersecurity concerns due to its potential for exploitation.
What are the cybersecurity risks associated with GLM-5.3?
The report by Anthropic's Frontier Red Team indicates that GLM-5.3 can be exploited by anyone who downloads and modifies it, with a success rate of generating working exploits in about 12% of attempts.
How does GLM-5.3's safety protocols compare to other models?
Anthropic's report highlights that GLM-5.3 has several methods to bypass its safety protocols, raising questions about the effectiveness of safety measures in open-weight AI systems.
What is the cost of stripping safety features from GLM-5.3?
The cost of removing safety features from GLM-5.3 is approximately $4,400, while still retaining most of the model's functionality.
What is Anthropic's stance on AI safety testing?
Anthropic CEO Dario Amodei supports targeted safety testing for advanced AI models, regardless of whether they are open or closed, rather than advocating for outright bans.