Researchers at Epoch AI have demonstrated that Chinese open-weight AI model GLM 5.3 can autonomously create working Firefox exploit chains without human intervention. By targeting the “patch gap”—the timeframe between public code fix commits and official user rollout—the model successfully generated multi-stage zero-day exploits across six Firefox releases at an average API cost of $1,000 per vulnerability.

Operating in an offline environment with standard developer tools, GLM 5.3 autonomously identified security-related code changes, bypassed sandbox protections, and built web pages capable of exfiltrating local system files. Epoch AI reported that public API providers triggered no safety guardrails during the experiment. Similar automated attack setups against Google Chrome required $100,000 in API costs without producing a viable exploit.

The findings were disclosed to Mozilla, which resolved the underlying sandbox vulnerability in Firefox version 159.0. The study highlights growing concerns regarding the offensive cyber capabilities of accessible open-weight models, underscoring the urgent need for software vendors to shorten patch deployment cycles across production software.

Why it matters

  • Autonomous AI exploit generation fundamentally shortens the safe window software vendors have to distribute security patches to users.

  • API providers and model developers face rising pressure to deploy effective guardrails against automated offensive cybersecurity workflows.

  • Enterprise security teams must prepare for low-cost, AI-driven zero-day exploitation targeting unpatched client software.

Source: the-decoder.com