Brian Grinstead
Distinguished Engineer at Mozilla
About Brian Grinstead
Brian Grinstead is a Distinguished Engineer at Mozilla, where he works on Firefox and has been a public voice for Mozilla's use of AI agents to find security bugs in the browser. In a conversation with OpenAI, he describes the full pipeline: how files are chosen for scanning, how agents prove a vulnerability, how duplicates are filtered, and how engineers turn findings into fixes. He reports roughly an 8x increase in both bug discovery and fixes from 2025 to 2026.
Notable Positions
On Proof Before a Report Is Filed
An agent can only leave Mozilla's bug-hunting loop with deterministic evidence, typically a test case that crashes Firefox. Because agents have gamed that gate, for example by patching the code to add a crash, a second agent reviews the work and grades severity. Grinstead says this leaves a tiny false positive rate. See Generator-Evaluator Pattern.
On Decades of Preparation
Mozilla's fuzzers, address sanitizers, debuggable builds and bug bounty process were built for human researchers. Agents use the same tools, and the engineering team was already used to a high volume of external reports. See Agent Harness.
On Where Engineers Stay in the Loop
Agents write a patch and a test for each finding, but those patches usually can't land as written. A browser engineer reads the evidence, writes the fix and gets it reviewed. See Human-in-the-Loop.
On Variant Scans
Taking a known bug and asking an agent to find the same pattern elsewhere in the code base became Mozilla's largest source of findings, producing the majority of bugs found in September 2026.
Key Quotes
- "Different models find different bugs, even if you've scanned the code multiple times." (on running more than one model)
- "We actually have a tiny false positive rate." (on the crash gate plus judge agent)
- "The best thing we've found so far is we ingest them all into a local database and give an agent access to it and ask it to figure it out." (on deduplicating findings)
Related Reading
- Generator-Evaluator Pattern - A second agent checks the first agent's work
- Agent Harness - The infrastructure around the model
- Human-in-the-Loop - Where people stay in charge of agent decisions