A Researcher at OpenAI's Codex Security Got Named Credit in a Chrome Advisory, Then in Firefox's, Same Day
Google shipped Chrome 151.0.7922.169/.170 on August 18, patching 15 security vulnerabilities including two rated Critical. One of the 15, a High-severity use-after-free bug in WebGL, carries a credit that no prior Chrome stable advisory has: Amy Burnett, a researcher working in OpenAI's Codex Security division. The same day, Mozilla's Firefox 154 advisory credited her for five more high-severity bugs. One researcher, using agentic security tooling built by an AI lab, landed confirmed findings in two competing browser codebases on the same patch day. That's a more interesting and more precise story than "AI found a Chrome bug," and it's worth being precise about what actually happened versus what a faster read of the headlines implies.
What actually shipped, and what didn't come from OpenAI
The two Critical bugs in this release are CVE-2026-76034, a buffer overflow in WebGL, and CVE-2026-76036, a buffer overflow in Dawn, Chrome's WebGPU implementation. Both were reported by Google's own team (July 15 and July 28 respectively), not by OpenAI. Both are rated Critical specifically because they live in Chrome's GPU process rather than its renderer: the renderer sandbox is what normally contains a compromised web page, but the GPU process sits outside that boundary by design, so a buffer overflow there doesn't need a separate sandbox-escape step to reach the operating system.
The bug Amy Burnett is credited for, CVE-2026-76045, is a separate, High-severity use-after-free in WebGL, reported August 5. It's a real, confirmed, patched vulnerability, and being named in a Chrome stable advisory at all puts it in a minority: of the 15 total fixes in this release, 12 came from Google's own internal tooling, and only three, including Burnett's, came from outside researchers. But it is not one of the two Critical, sandbox-adjacent bugs, and getting that distinction right matters more than the headline version of this story.
The part that's actually new
Burnett works in OpenAI's Codex Security division, an application-security agent OpenAI launched in March 2026 that builds a threat model of a codebase before scanning it for vulnerabilities; OpenAI has claimed the tool surfaced roughly 800 critical and over 10,000 high-severity issues across projects including Chromium, OpenSSL, and PHP during a 30-day window after launch, a vendor-reported figure worth treating as a claim rather than an independently audited count. What's independently verifiable is narrower and, in some ways, more interesting: her CVE-2026-76045 credit in Chrome's advisory landed the same day Mozilla's MFSA 2026-74 credited her for five separate high-severity Firefox bugs. According to her published CyberUK 2026 speaker bio, Burnett has over a decade of vulnerability research experience and previously took part in DARPA's AI Cyber Challenge. This isn't a model operating unsupervised and getting a byline. It's a human security researcher, using agentic tooling built at an AI company, credited by name for confirmed findings in two separate, competing, high-scrutiny browser codebases inside the same 24 hours.
That's a meaningfully different and more defensible claim than "an AI found a browser zero-day." It's evidence that agentic security tooling, in the hands of an experienced researcher, is now productive enough to generate confirmed, patched findings across multiple large codebases in parallel, fast enough to land in two vendors' advisories on the same day. Worth naming directly: Google's own comparison point undercuts any narrative that this represents AI security tooling overtaking human researchers broadly. Google's internal Gemini-powered pipeline is credited with fixing 1,072 vulnerabilities across the Chrome 149 and 150 release cycles combined, dwarfing the volume of any single external researcher, AI-assisted or not. External contributions, Burnett's included, remain a small minority of any given Chrome release. What changed isn't the ratio, it's that one specific, named, AI-tooled researcher just demonstrated cross-codebase reach in a single day, which is a new kind of data point even if it's a small one.
What this means for a solo operator today
Nothing changes about your Monday because of this specific release. You're not running Codex Security, the two bugs that actually matter for your own browser security hygiene (the Critical WebGL and Dawn overflows) came from Google's own team, and the honest takeaway isn't "adopt AI security tooling now." I've read enough of these AI-tool-finds-a-bug stories this year to know the pattern: the headline outruns what the finding actually supports, and the useful reaction is to note the trend line rather than change anything today. What's actually worth tracking is that an individual researcher armed with agentic security tooling generating confirmed, cross-codebase findings on the same day is a preview of what a much larger population of researchers, and eventually tools available directly to smaller teams, could look like as this category matures. When something like Codex Security, or a comparable product, ships as a tool I could point directly at this site's own codebase rather than at browser engines and other researchers' target lists, the questions that will actually matter are unglamorous: what's the false-positive rate on the actual codebase, what does it cost against a human audit, and does it find the bug classes that threaten a small SaaS (auth bypass, IDOR, leaked secrets) rather than only the memory-corruption classes that threaten browser engines specifically.
The honest counter-take
It's easy to overread a same-day, two-browser credit as a bigger signal than it is. Twelve of Chrome 151's 15 fixes still came from Google's own internal fuzzing and testing infrastructure, external researchers as a category remain a small minority of any release, and one experienced human researcher using capable tooling is a different and much less dramatic claim than "AI models are now finding browser zero-days on their own." The genuinely new thing here is narrower than the headlines suggest: agentic tooling in skilled hands is fast and portable across codebases. That's worth watching. It is not yet evidence that this replaces, rather than augments, the security researchers who already do this work.
Author
Lukas
@lukcombinator