According to a report by The Decoder, researchers at Epoch AI have found that the open-weight model GLM 5.3 autonomously developed full exploit chains against the Firefox browser without any human intervention in less than 48 hours. The widely available AI model utilized public security fixes to create working exploits before updates could reach end users. The automated attacks successfully compromised six different Firefox releases, with each complete exploit chain costing approximately 1,000 dollars in API fees. The evaluation also noted that the process bypassed the safety safeguards of every provider that was tested.
Implications for Automated Vulnerability Discovery
The capability of open-weight models to autonomously weaponize public security patches represents a significant shift in software security dynamics. Traditionally, discovering zero-day vulnerabilities or turning patch diffs into functional exploits required specialized human engineering effort and substantial time. When an AI model can perform this reconnaissance and exploit construction automatically and at scale, the window of vulnerability between a patch being published and users applying the update narrows dangerously. This development highlights growing concerns among researchers regarding the dual-use nature of capable AI models, particularly as API access becomes cheaper and safety guardrails fail to prevent malicious utilization.
Unanswered Questions and Future Monitoring
Several critical details regarding the incident remain unknown. The announcement did not specify which exact versions of Firefox were targeted beyond the six releases mentioned, nor did it disclose the identities of the specific API providers whose safeguards failed to trigger during the tests. Furthermore, public documentation has not revealed the precise prompt engineering or workflow orchestration used to guide GLM 5.3 through the multi-step exploit generation process, leaving the exact methodology opaque to the broader cybersecurity community.
As open-weight models continue to advance, researchers and security professionals will need to closely monitor how rapidly AI-driven vulnerability exploitation outpaces traditional defensive patching timelines.
Lead image: Tima Miroshnichenko / Pexels
