News · 2026-07-30

Google cut Chrome's bug bounty payouts because its own AI now finds too many bugs

Google has changed how much it pays outside researchers for Chrome security bugs because its own AI tooling now finds so many of them. In its quarterly Chrome Security update for the second quarter of 2026, the team writes that the Chrome vulnerability reward program "adjusted its reward structure and reward amounts to reflect the volume of reports being discovered and fixed using internal AI tooling." The same update says Google's Big Sleep agent now runs as a fully automated pipeline against V8, Chrome's JavaScript engine.

Key facts

• Google's Chrome VRP changed both its reward structure and reward amounts because of AI-discovered bug volume.
• The Big Sleep agent "now operates as a fully automated pipeline, helping secure V8."
• Chrome is "moving to more frequently rejecting reports" that do not meet reporting guidelines, and is reworking CVE issuance to scale with the volume.
• Primary source: the Chrome Security 2026 Q2 update, posted by Jasika Bawa on behalf of Chrome Security.

Bug bounties are a market. A browser vendor pays outside researchers for vulnerabilities because outside researchers are cheaper and more numerous than in-house ones, and because a bug bought is a bug not sold to someone worse. That market prices scarcity. When the vendor's own tooling starts producing the goods in bulk, the price moves.

That is what Google is describing, in the flattest possible language, buried in a routine quarterly newsletter. The full sentence is worth reading twice: the reward program changed "to reflect the volume of reports being discovered and fixed using internal AI tooling." Google is not saying AI helps its researchers. It is saying AI output is now large enough to reprice an external market that has run since 2010.

The mechanism behind the automation is the interesting part, and it is mundane engineering rather than a leap in model capability. The blocker for any autonomous bug-finder is not finding candidate crashes - it is that most candidates are noise, and a human has to sort them. Google's update explains that V8 "now provides better modes for determining where crashes happen," with new testing modes that distinguish experimental from production issues. That, the team writes, "allows our internal bug finding agents to automatically validate all discoveries, ensuring that they only report real bugs. As a result, the Big Sleep agent now operates as a fully automated pipeline."

In other words, the unlock was building a reliable oracle for "is this real?" The agent did not get smarter; it got a grader. That is the same shape as the verifiable-reward pattern that has driven so much recent progress - automation becomes possible exactly when checking is cheap and certain.

The scale is visible elsewhere in Google's own numbers. Google DeepMind reported in July that its lightweight cyber model found 55 unique confirmed issues in V8 in a production commit-scanning evaluation, beating both a larger general model and a rival frontier model. The Chrome team now says it is triaging with a mix of deterministic and AI tooling on isolated infrastructure that reproduces and enriches reports before routing them, and that there is "ongoing work to also improve our CVE and release notes issuance processes to scale appropriately with the increased volume we are observing." When your bottleneck becomes issuing CVE identifiers fast enough, the finding problem has changed character.

The counterweight landed the same week from an entirely different corner of open source. The LLVM compiler project's AI Tool Use Policy, adopted in January 2026, permits AI assistance but insists that "there must be a human in the loop." Contributors must read and review anything a model generates, remain accountable as the author, be able to answer questions about their work in review, and label substantial machine-generated content. Agents that act in LLVM's spaces without human approval are banned outright, as are review bots that post without human review - and using AI tools to fix issues tagged for newcomers is forbidden, to protect the on-ramp for new contributors.

So two of the largest codebases in open source are pulling in opposite directions on the same question, and both are being honest about why. Chrome is industrialising machine-found bugs and repricing the humans around them. LLVM is drawing a line at machine-submitted work and defending the review culture. Neither is obviously wrong; they are optimising different things.

The honest caveat is about what "fully automated" covers. The claim is scoped to discovery and validation in V8, not to writing and landing patches across Chrome. Google's separate work on CodeMender, the agent that proposes fixes rather than finds bugs, still states that "all patches generated by CodeMender are reviewed by human researchers before they're submitted upstream." A widely circulated version of this story claims Google's AI fixed more Chrome bugs in June than in the previous two years combined; that specific comparison does not appear in any Google source we could find, and the quarterly update publishes no June count and no two-year baseline at all.

Primary source, verified: read the paper →

Ground Truth. (2026, July 30). Google cut Chrome's bug bounty payouts because its own AI now finds too many bugs. Ground Truth. https://groundtruth.day/news/google-cut-chrome-bug-bounties-because-its-own-ai-finds-too-many-bugs.html

@misc{groundtruth:google-cut-chrome-bug-bounties-because-its-own-ai-finds-too-many-bugs,
  title  = {Google cut Chrome's bug bounty payouts because its own AI now finds too many bugs},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/google-cut-chrome-bug-bounties-because-its-own-ai-finds-too-many-bugs.html}
}

Topics: cybersecurity · ai-security · google · vulnerabilities · chrome · agents · bug-bounty