According to a report by The Decoder, artificial intelligence firm Anthropic has stated that Zhipu‘s open-weight model, GLM-5.3, approaches the capability of Claude Mythos Preview when generating cyber exploits. Additionally, a smaller variant named Flash reportedly developed a functioning Chrome attack for a cost of twenty dollars and forty cents based on Zhipu API pricing. Anthropic noted that the model’s safety guardrails can be bypassed easily, with unlocked versions of the software already spreading publicly. The United States agency CAISI has reportedly corroborated these findings.
The Open-Weight Safety Dilemma
The development highlights ongoing industry concerns regarding open-weight artificial intelligence models. When model weights are made accessible to the public, safety features implemented by developers can often be stripped away by users. This dynamic lowers the technical barrier for malicious actors to utilize advanced capabilities for cyber attacks, creating challenges for regulators and developers alike.
Unanswered Questions
Several details regarding the evaluation remain unknown. The announcement did not specify the exact methodology used to measure exploit capabilities, nor did it disclose the full scope of the testing performed by CAISI or Anthropic. Furthermore, the long-term impact on enterprise security and the specific steps Zhipu plans to take regarding the circulating unlocked versions were not detailed in the available reports.
Observers will need to monitor how regulatory bodies and AI labs respond to the proliferation of capable open-weight models with easily removable safeguards.
Lead image: Brett Sayles / Pexels
