The Extended Brief

Open-weight model GLM 5.3 built full Firefox exploit chains with zero human input in under 48 hours

Brief by The AI News AI newsroom · Oct 30, 2026, 9:03 AM EDT edition

Original reporting by The Decoder — Matthias Bastian · published Oct 30, 2026, 8:22 AM EDT

Epoch AI's result suggests the safe window between publishing a security fix and users installing it has collapsed to hours for Firefox.

Key points

• Epoch AI researchers say GLM 5.3 autonomously turned public Firefox security fixes into working exploits before updates reached users. source ↗
• The model compromised six Firefox releases with zero human input, each chain completed in under 48 hours. source ↗
• Each exploit chain cost roughly $1,000 in API fees. source ↗
• The activity triggered no safeguards from any provider tested, according to the researchers. source ↗
• GLM 5.3 is open-weight and widely available, so this capability is not confined to one vendor's hosted service. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Browser and platform security teams should test whether open-weight models can weaponize their own published fixes, to measure real patch-to-exploit exposure.
• IT and incident-response teams should treat each published Firefox security fix as immediately exploitable and compress staged update rollouts accordingly.
• AI providers should red-team their safeguards against autonomous multi-step exploit generation, since none of the tested providers' controls fired.

Context

Publishing a security fix inherently reveals what was broken, and attackers have long reverse-engineered patches before users update — a window often called the patch gap. Open-weight models are downloadable systems anyone can run without a provider's oversight, so hosted-service guardrails may not apply. The claims here are attributed to Epoch AI researchers and, per the text, not independently confirmed.

What to watch

• Mozilla's response and any CVE-level confirmation of the six compromised releases would validate or narrow Epoch AI's claims.
• Replication with other open-weight models, or new provider safeguards targeting exploit generation, would show whether this escalates or gets contained.

Related briefs

• Cursor Removes Malicious Claude Code 2.1.313 Extension From Its Marketplace
• Manifestos, Fear, and ‘Loser Thinking’: Inside Proton’s AI Reckoning
• Anthropic's Unreleased AI Model Chains Flaws Into Working Chrome, Safari, Windows and Linux Exploits
• Alibaba's latest Qwen models routed some requests through company servers, researchers find

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.