AI patch gap exploit: Models now write code to attack Firefox

Epoch AI researchers recently demonstrated a successful AI patch gap exploit that turns public security fixes into immediate threats. Using the open-weight model GLM 5.3, they showed the AI could autonomously develop full exploit chains for Firefox. According to The Decoder and shattered.io, the model successfully targeted six Firefox versions released between September 17 and October 26.

The attack leverages the window between a developer publishing a code fix and users actually installing the update. Real Bytes and the Mozilla Blog note that AI can now compress this window to under 24 hours. The Cloud Security Alliance considers this a permanent structural shift, fundamentally advantaging attackers by dropping the cost of exploiting these vulnerabilities to near zero.

These systems do not just write malicious payloads. Real Bytes explains that models rapidly analyze the code differences between vulnerable and patched versions to map out exact exploit vectors. Horizon3.ai adds that an AI can run thousands of hypothesis tests in the time a human researcher finishes a few dozen.

The efficiency is startling. The Epoch AI demonstration took less than 48 hours to create a working exploit chain. It cost approximately $1,000 in API fees per successful attack. If you want to estimate the financial overhead of running thousands of prompts for similar automated tasks, you can compare model pricing with our LLM Cost Calculator.

SecureLayer7 points out that AI vulnerability research is a dual-use capability. Defenders can use these models to triage and fix flaws faster, but attackers use the exact same capability to weaponize publicly available patches. The Google Threat Intelligence Group recently identified a threat actor using a zero-day exploit believed to have been developed with AI, confirming that models are actively chaining vulnerabilities in the wild.

Developers also need to watch their own generative coding assistants. ITU Online IT Training reports that tools like GitHub Copilot or Cursor AI can inadvertently introduce new vulnerabilities into custom applications by generating insecure code with known flaws. If you are feeding large blocks of code into an assistant to check for these flaws, use a token counter to ensure you stay within the model's context limits.

The traditional grace period of days or weeks to apply a security patch no longer exists. With models turning known vulnerabilities into working exploits in hours, software vendors have to accelerate their deployment pipelines.

For IT teams, the immediate takeaway is strict, automated patch management. If an update is public, assume an exploit is already being generated. You cannot outpace an automated system manually, so your deployment systems need to be as fast as the tools analyzing your code.