# How Mozilla uses GPT-6 Cyber to find and fix Firefox vulnerabilities

## Executive summary

Mozilla is integrating advanced AI agents, such as GPT-6 Cyber, into its security workflow to significantly scale vulnerability discovery and remediation for Firefox. The process involves an 'agentic security pipeline' that uses LLMs for file prioritization, runs bugs through a controlled inner loop (fuzzing and static analysis), and employs specialized agents for deterministic verification and patch generation. This approach has led to an observed 8x increase in bug discovery and fixes between 2025 and 2026.

## Key takeaways

- Increased Bug Discovery Rate: Mozilla reported an 8x increase in both bug discovery and fixes from 2025 to 2026, driven by the new agentic security pipeline.
- Variant and Commit Scanning: Variant scanning, which searches for known bug patterns across the entire codebase, is proving to be the majority source of newly found bugs, making it a key part of the current strategy. Commit scanning is also being implemented for narrow-scope vulnerability checks.
- AI-Assisted Prioritization: LLMs are used to score files based on risk (e.g., exposure to web page content, memory safety characteristics) to guide the agent on where to focus scanning efforts, moving beyond simple alphabetical file traversal.

## Technical details

- Agentic Security Pipeline Architecture: The pipeline includes an outer loop for bug discovery and an inner loop for testing. The inner loop places an agent in a VM with access to the Firefox source code and fuzzing toolchain. The agent generates bug theories (often via static analysis) and tests them dynamically until deterministic verification (e.g., an HTML file causing a crash) is achieved.
- Verification and Remediation: After a bug is found, a separate agent performs deterministic verification and severity grading. This process reduces the false positive rate. Furthermore, an agent can write code patches and accompanying test cases to support the engineering team in fixing the vulnerability.
- Vulnerability Deduplication: To manage the high volume of findings, Mozilla ingests all signals into a local database and uses an agent to help identify duplicates, recognizing that similar crash stacks can represent different bugs.
- Agentic Vulnerability Workbench: A centralized workbench ingests all signals (bug reports, comments, runs) into a single database. This allows engineers to query the data using a coding agent to plan future scans or diagnose issues with the harness itself (e.g., identifying missing system dependencies).

## Practical implications

- Security teams can leverage LLMs to move beyond manual prioritization, using risk scoring and historical data to guide scanning efforts.
- The integration of automated patch generation and test case creation significantly accelerates the remediation phase of the security lifecycle.
- The use of variant scanning suggests a shift toward pattern-based vulnerability hunting rather than solely relying on known exploit vectors.

## Topics

Browser Security, AI Agents, Vulnerability Research, Fuzzing, Code Analysis, Software Reliability, Firefox, GPT-6 Cyber

Source: https://www.youtube.com/watch?v=PkAmyub12K8
