Google Paused Its Open-Source Bug Bounty Over AI Slop -- Nine Months After curl Hit the Same Wall

Google stopped accepting new reports to its Open Source Software Vulnerability Rewards Program on Oct. 1, after what the company called "a significant rise in automated submissions, the vast majority of which are not valid." The program, running since 2022, covers Google's own open-source projects -- Go, Angular, Bazel, Protocol Buffers, Fuchsia -- plus the repository settings and supply-chain components around them, with per-bug rewards historically ranging from $100 to $31,337 depending on severity and project tier. Reports filed before Oct. 1 are unaffected, and Google says it will give "an update" in the first quarter of 2027 -- not a return date, an update.

Google didn't publish a submission count or a validity rate for its own program. But the shape of the problem has a fully documented, nine-months-earlier precedent. curl -- the command-line tool and library embedded in nearly every piece of networked software -- ended its own HackerOne bounty on Jan. 21 for exactly the reason Google now cites. Maintainer Daniel Stenberg wrote that roughly a fifth of reports were already AI-generated, and that only about 5% of all submissions, AI-written or not, described a real vulnerability.

The main goal with shutting down the bounty was to remove the incentive for people to submit crap and non-well researched reports, AI generated or not.

The problem isn't confined to curl or Google. A Cloud Security Alliance research note published this month puts numbers to what security teams across the industry have been describing anecdotally for most of the year: HackerOne logged a 210% year-over-year rise in AI-related reports, and Bugcrowd's submission queue spiked 334% over a single three-week stretch. Not all of it is noise -- HackerOne's own figures credit fully autonomous AI systems with more than 560 valid findings over the past year -- but distinguishing those from the flood is exactly the labor-intensive work bounty programs are now buckling under.

The strain isn't contained to bounty programs, either. The CVE database recorded 48,185 new entries in 2025, a 20.6% jump over 2024, while the National Vulnerability Database's enrichment -- the analysis that tells a defender how serious a given CVE actually is -- covered only 28% of that year's entries, down from 46.2% the year before. The CSA note's recommendation is blunt: stop paying for low-value findings at all, require sandboxed proof of exploitability before a report is even reviewed, and treat the enrichment gap as a structural risk rather than a backlog that will eventually clear.

The economics are structural, not a one-time spike. Generating a plausible-sounding vulnerability report with an AI coding assistant now takes seconds; confirming or rejecting one still takes a human security engineer roughly the same time it always has. Malwarebytes reported that Intel made a quieter version of the same move this year, narrowing what it pays bounties for on the Intigriti platform rather than shutting a program down outright -- a sign the asymmetry isn't unique to a volunteer-run project like curl or to one company's choices.

The size mismatch is what makes Google's move notable rather than simply predictable. curl's entire bounty history, since 2019, paid out just over $100,000 across 87 confirmed vulnerabilities. Google's 2025 alone -- across every one of its vulnerability reward programs, not the OSS VRP specifically -- paid a record $17.1 million to more than 700 researchers, up 40% from 2024's $12 million, part of $81.6 million paid out since 2010. A rewards economy roughly two orders of magnitude larger than curl's broke under the same kind of load, which suggests the volume of low-effort AI submissions scales with how attractive a program's payouts look, not with how large or well-resourced the defender is.

Google's own framing points toward a fix that filters rather than reopens -- using AI to pre-screen submissions before a human reviewer ever sees them, the approach several of this week's reports describe the company weighing. HackerOne has already shipped something similar for its own clients, called h1 Validation, combining automated analysis with human review to separate exploitable findings from the rest. That would turn the tool that caused the problem into the tool that triages it, which is plausible but untested at Google's scale, and not something Google has committed to in writing. Until the Q1 2027 update arrives, the honest status of Google's open-source bug bounty is paused, not fixed -- and every other maintainer running a program large enough to be worth gaming is watching to see what that update actually says.