AI can turn security fixes into working exploits before users receive an update
Researchers found that the open-weight AI model GLM 5.3 can autonomously exploit the "patch gap"—the window between a security fix being published and reaching users—to create working exploits for recent versions of Mozilla Firefox. Operating with minimal resources and no internet access, the model built a website capable of reading local files without user interaction by chaining a patch-gap vulnerability with an unpublished flaw. Open article →
Notable comments
Discussion broadly converges on patch-gap risk symmetry from open-weight models while stressing that the practical lever is speeding/stratifying fixes to reduce exploitable windows, not banning AI.
HN · mozdeco Mozilla author account points readers to a concrete defense approach for handling the patch gap, offering primary-source context beyond the researchers’ claim.
HN · pizza234 Pushes beyond generic 'AI helps attackers' by arguing attacker capability costs (e.g., hundreds of GB RAM for GLM 5.3) vs defense diversity create an asymmetric near-term risk landscape.
HN · mozdeco Provides a specific mitigation strategy: avoid shipping patches where renderer vulnerabilities and sandbox escapes land in the same patch-gap window, and notes why fully closing the gap is infeasible.
- Software updates aren’t what they used to be
- Open-weight model GLM 5.3 built full Firefox exploit chains with zero human input in under 48 hours
- Epoch AI showed that GLM 5.3 autonomously reverse-engineered published but not yet shipped Firefo...
- HOLY MOLY: Epoch AI spent over $100,000 testing https://t.co/HLWjyfsvyK's open-weight GLM 5.3 aga...
- Epoch AI researcher @AlexBarry4 warns that banks, hospitals, and water systems may not be able to...
- Epoch AI researcher @AlexBarry4 warns that ignoring software updates is becoming far more dangero...
- Epoch AI researcher @AlexBarry4 reveals how AI can turn a company's own security fixes into worki...
- Epoch AI researcher @AlexBarry4 reveals how a $1,000 AI-generated attack could steal private file...
- GLM-5.3 can exploit the period between a security fix becoming publicly available and the fix rea...
- Open-weight AI models can exploit the patch gap
- GLM 5.3 can autonomously compromise recent versions of Mozilla Firefox
- as best as i can tell, this is a pretty general inditement of the security properties of using *a...