GLM 5.3 created exploits for Firefox without human intervention

Researchers at Epoch AI reported on the autonomous operation of the GLM 5.3 model, which was able to turn public security patches into functional exploit chains. The process occurred without human intervention.

The artificial intelligence compromised six releases of the Firefox browser before the updates could reach regular users. Each attack took less than 48 hours.

The cost of API requests was approximately $1,000 for each successful incident. Furthermore, none of the tested providers recorded any triggers of their built-in safety guardrails.

The results demonstrate new cybersecurity risks associated with the autonomous use of open models to generate complex vulnerabilities.