Learn about this first
An experiment by Epoch AI showed that GLM 5.3, an open-weight model released by Z.ai in August, could autonomously construct web attacks without human input by exploiting security patches that Mozilla had released but not yet delivered to users (a patch gap). Visitors could have their computer files read simply by visiting a page, and all six Firefox versions released between September 17 and October 26 were compromised.
The experiment gave the model up to 48 hours to find exploits; its fastest attempt took only 12 hours. Each successful attack cost approximately $1,000 in API fees. Epoch AI said that all attacks against the latest Firefox versions required chaining a patch-gap vulnerability with an undisclosed vulnerability to bypass standard defenses.
The Decoder reported that when researchers used GLM 5.3 to develop exploits, neither Z.ai nor other API providers triggered safety guardrails or refused the requests.
AI-generated from reports · updated 14 days agoLatest developments
Epoch AI test: Open-source model GLM 5.3 builds a working attack by exploiting an unreleased Firefox security patch
Event progress1
Epoch AI publishes research on using GLM 5.3 to exploit patch gaps to attack Firefox
Epoch AI: Gradient Updates: Epoch AI experiment shows that GLM 5.3 can autonomously attack Firefox by exploiting patch windows