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

What happened

Epoch AI researchers demonstrated that the open-weight AI model GLM 5.3 automatically generated full exploit chains for Firefox vulnerabilities without any human prompts. In less than 48 hours, the model turned recent public security fixes into operational exploits affecting six Firefox versions. Each exploit cost around $1,000 in API usage fees and successfully bypassed protection systems from every tested AI provider.

The risk

This shows that current AI models can autonomously weaponize browser vulnerabilities faster than patches reach users. The speed and independence of GLM 5.3 in building exploits cuts dramatically into the traditional security window between public disclosure and widespread exploitation. It also reveals a blind spot in existing AI safety systems, which failed to prevent the generation of working browser exploits.

Why it matters

For defenders, this shifts the security calculus. Vendors and users can no longer rely on human-led exploit development slowing attack rollout. Patch deployment speeds now face pressure to accelerate sharply to stay ahead of automated AI-driven attackers. The cost barrier to exploit development drops too, enabling more adversaries to fund attacks cheaply through public APIs. This raises the stakes for browser security teams, threat intelligence, and API providers, which must develop new detection and mitigation tactics for autonomous exploit generation.

Who should pay attention

Security teams at browser vendors, especially Firefox, have urgent reason to tighten their update processes and monitoring for AI-driven attack vectors. API providers supplying AI models need stronger safeguards to detect and block exploit chain generation. Enterprises relying on Firefox should reassess risk exposure and patch strategy. AI developers and researchers monitoring model misuse must recognize that unrestricted open-weight models are a high-risk vector for rapid vulnerability weaponization.

What to watch next

Look for accelerated patch release cadence from Firefox and other browsers aiming to close the public window before automated exploits arrive. Observe whether AI providers strengthen their usage filters or throttle access to code generation for security exploits. Expect increased regulatory and industry pressure on both AI vendors and software suppliers to manage this emergent risk. Threat intelligence feeds will need new signals for identifying AI-crafted attacks moving from theory to real-world impact.

AI Quick Briefs Editorial Desk