Anthropic put out a report that deserves more attention than it will probably get. Their security team tested an open model out of China called GLM-5.3, built by a company named Zhipu, also known as Z.ai. They wanted to know how good it had gotten at building real cyberattacks without any help.
The results are uncomfortable. In testing, GLM-5.3 built a working exploit from scratch in about twelve percent of attempts, close to what Anthropic’s own strongest model can do. A year earlier, models in this same class scored close to zero on those exact tests.
One detail stuck with me. A cheaper, faster version of the model turned two known bugs in Chrome into a working attack chain in about eight hours. Roughly twenty dollars and an afternoon, nothing close to what a real operation used to cost a nation state to pull off.
The safeguards built to stop this kind of misuse barely held up in testing. Researchers got around them most of the time using simple tricks, nothing close to advanced hacking skill.
I am not writing this to scare anyone off AI. The same tools that can write an exploit can help a defender find the hole first and patch it before anyone else looks. Anthropic makes that point themselves in the report. But when an open model can build attacks this fast, with safeguards this thin, the people defending networks need to move at the same speed. Take it seriously, stay alert, and keep building anyway. That is the only response that actually holds up. Open weights mean anyone can download this thing and run it on their own machine tonight, no permission needed, so the clock on fixing the gap is already running.
——
Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277
——
Sources:
https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities
https://www.trendingtopics.eu/anthropic-glm-5-3-cyber-warning/
https://aiweekly.co/alerts/anthropic-glm-53-marks-step-change-in-attacker-cyber-tools
Repost this. Thanks.
