FULL STORY
GLM 5.3 Autonomously Exploits Six Firefox Versions for ~$1,000
Epoch AI reported that open-weights model GLM 5.3 can autonomously exploit six Firefox versions for about $1,000 using patch-gap N-days; follow-up tests a week later confirmed the findings.
2026-10-30 ~ 2026-11-07 · 3 episodes · 12 posts
Episode 1 · Epoch AI Experiment: GLM 5.3 Autonomously Exploits Patch Gaps to Break Six Firefox Versions (2026-10-30, 6 posts)
An experiment by Epoch AI researcher Alex Barry's team showed that the open-weight model GLM 5.3 can autonomously compromise multiple recent versions of Mozilla Firefox by exploiting the "patch gap"—the window in which security fixes are already public in development repositories but have not yet been pushed to stable users, a window the experiment describes as potentially just a few hours.
Confirmed
- Given only a simple initial instruction (to check fixes that had not yet shipped), GLM 5.3 autonomously identified Mozilla's already-fixed but undelivered security patches and reverse-engineered them into working exploits, without deep human involvement.
- Using only information publicly available at each release, the model broke six recent Firefox versions, including versions released in September.
- According to @Elivnurotic and @Jsevillamol, the cost per exploit was only about $1000.
- Per an excerpt relayed by Kelsey Tuoc via @austinc3301, the model could also build malicious web pages that exploit Firefox users on visit without further interaction.
- The experimental setup deliberately simulated a resource-limited average attacker rather than a nation-state advanced threat actor.
Why it matters
- The results show that frontier open-weight models already possess end-to-end autonomous vulnerability discovery and exploit-building capability at far lower cost than traditional exploit development, sharply lowering the barrier.
- The patch gap is already a real weakness of browsers and similar software; maturing AI capability makes attacks within the "fix published—user updated" window far more feasible, raising the bar for vendors' patch release and rollout practices.
- Epoch AI: GLM 5.3 autonomously exploits unpatched flaws to compromise six Firefox versions — Miles_Brundage · 2026-10-30
- Epoch AI: GLM 5.3 autonomously builds Firefox exploits from the patch gap for ~$1,000 per version — Eliv_nurotic · 2026-10-30
- Epoch AI: GLM 5.3 autonomously exploits Firefox patch gap for ~$1,000 per version — Jsevillamol · 2026-10-30
- Epoch finds GLM autonomously exploits unshipped Firefox fixes for ~$1,000 each — Jsevillamol · 2026-10-30
- Epoch AI: GLM autonomously exploits Firefox using unshipped security fixes at ~$1,000 per version — Jsevillamol · 2026-10-31
- Open-weight models exploit Firefox patch gaps for ~$1,000 per version — austinc3301 · 2026-11-01
Episode 2 · Open-source GLM 5.3 Exploits Patch Gaps to Break Six Firefox Versions (2026-11-02, 2 posts)
Epoch AI reported that the open-weight model GLM 5.3 can exploit patch gaps to compromise six Firefox versions at a cost of about $1,000 per attempt, highlighting security risks from delayed patch rollout.
- Epoch AI: open-weight GLM 5.3 exploited the patch gap to hack 6 Firefox releases at ~$1,000 each — rohanpaul_ai · 2026-11-02
- Epoch AI report: open-weight models can exploit the patch gap, 6 Firefox releases compromised — rohanpaul_ai · 2026-11-02
Episode 3 · Epoch: GLM 5.3 reproduced six Firefox n-day exploits for ~$1000 (2026-11-07, 4 posts)
Epoch AI found GLM 5.3 could build working exploits for six Firefox n-days for about $1000 with no provider intervention; Matthew Green noted the uncensored variant refuses far less, calling unrestricted open models the real new development.
- Epoch test: GLM built working Firefox n-day exploits for ~$1,000 each, no refusals — rocklambros · 2026-11-07
- Epoch test: GLM 5.3 built working Firefox N-day attacks for ~$1,000 each, no provider refused — matthew_d_green · 2026-11-07
- Researcher says stock GLM 5.3 refuses often; abliterated version shipped same month at same price — matthew_d_green · 2026-11-07
- Security researcher: open models enabling full exploit builds is the real new capability — matthew_d_green · 2026-11-07