Open-Weight AI Matches Frontier Cyber Capabilities
According to an analysis by Anthropic cited by The Decoder, Zhipu’s open-weight GLM-5.3 model nearly matches the capabilities of the restricted Claude Mythos Preview in developing cyber exploits. The analysis states that GLM-5.3 can build complete cyber exploits on its own, marking a significant milestone for open-weight systems. Unlike comparable proprietary models, the software shipped without effective built-in safeguards that prevent malicious deployment.
Tests showed that GLM-5.3 successfully built working exploits and navigated complex security environments during benchmark evaluations. Anthropic noted that the model’s safety features could be bypassed using straightforward methods, including basic red-team framing and abliteration techniques to strip out refusal behaviors. The US agency CAISI reached similar conclusions in its independent assessment, labeling the model as the most cyber-capable open-weight system evaluated to date and highlighting potential risks for widespread misuse by state and non-state actors.
The findings arrive amid ongoing debates over the security implications of open-weight artificial intelligence. While developers and institutions praise open models for offering lower costs and customization options, security analysts warn that the rapid narrowing of capability gaps between proprietary and open systems presents persistent safety challenges. The UK AI Security Institute recently noted that open models have significantly reduced their performance lag behind leading US systems, complicating efforts to balance innovation with rigorous safety oversight.
Based on reporting by the-decoder.com.
