Epoch AI: open-weight GLM 5.3 built Firefox exploits for about $1,000 each

Epoch AI says it found that the open-weight model GLM 5.3 can autonomously compromise recent versions of Mozilla Firefox by exploiting the "patch gap," the window between when a security fix is published in a beta or nightly build and when it ships to regular users. Using a deliberately basic setup meant to resemble an unsophisticated attacker, Epoch gave the model only basic instructions to inspect not-yet-shipped fixes, and it reliably built working exploits for six recent Firefox versions on its own.

According to Epoch, these exploits cost about $1,000 per compromised version and would let an attacker steal stored passwords and authentication cookies or read files on a victim's computer after they open a single webpage, with no further interaction needed. Epoch gave the model up to 48 hours per attempt, though its fastest successful run took only 12 hours. In every successful attack, the patch gap gave the model its initial foothold; to go further and read files off the computer, the model also independently found and used a previously unpublished vulnerability.

Epoch shared its findings with Mozilla, which patched the newly found flaw in Firefox 159.0 on October 27. Epoch says it has not yet confirmed whether the attack still works without that vulnerability but expects the fix makes it substantially harder.

Why it matters

It suggests that exploiting the patch gap, once requiring real hacking skill, can now be done cheaply and quickly by widely available open-weight models.

Source: Epoch AI on X