AI Lowers Cyber Skill Barrier: Anthropic Study Shows Novices Can Launch Complex Attacks
- On Tuesday, Anthropic warned that GLM-5, an open-weight model from Chinese lab Zhipu, possesses powerful hacking abilities nearly matching its frontier model, Claude Mythos Preview, while lacking robust safety constraints.
- The San Francisco-based firm reported that GLM-5 safeguards can be bypassed in up to 100% of tests because the model's public weights allow developers to easily strip protections.
- Tests show the Chinese model completed 50 out of 410 exploit attempts, compared with 56 for Mythos Preview; GLM-5 achieved a full control-flow hijack in 4% of 100 tasks on internal benchmarks.
- Researchers used GLM-5 to discover unknown flaws in a JavaScript engine, chaining them into a webpage that reads files from a visitor's computer; Anthropic has disclosed the vulnerabilities to the maintainer.
- NIST's Center for AI Standards and Innovation labeled GLM-5 "the most cyber-capable open-weight model released to date," noting it lags the US frontier by about four months on aggregate cyber benchmarks.
28 Articles
28 Articles
The Zhipu AI model obtained results very close to Claude Mythos Preview in specific tests. Being an open source model (open weights), new debates arise about who can use these capabilities and what control their creators have.
The Standards and Innovation Center at AI also evaluated this model and concluded that it is “the one with the greatest capabilities in the field of cybersecurity launched to date.”
According to the researchers, attackers can bypass the protection of the model in 64-100% of cases using simple methods.
Anthropic said China's GLM-5.3 AI model can autonomously find vulnerabilities and create end-to-end exploits.
Antropic warned that while China's Z.ai's latest open-weight artificial intelligence (AI) model, 'GLM-5.3,' possesses cyberattack capabilities approaching those of its own top-tier model, its safeguards could be relatively easily neutralized. On the 1st, Hong
Coverage Details
Bias Distribution
- 50% of the sources are Center
Factuality
To view factuality data please Upgrade to Premium
