Google pauses open-source bug bounty program over AI-generated submissions

Google's own framing undercuts the easy narrative that AI-written reports are the problem by definition. An analysis of the pause argues the issue isn't that generative AI wrote the submissions - it's that so many of them are wrong. Poorly written, incomplete, and hallucinated reports are not always immediately recognizable as such ^[1], which is exactly what makes them expensive: a fabricated exploit path or a claim about a function that isn't reachable the way the report describes still has to be manually traced through source code before a reviewer can rule it out. Google confirmed the trigger was blunt volume, not just bad luck - a significant rise in automated submissions, the vast majority of which are not valid ^[2]. The two problems compound: more submissions mean more hours spent disproving plausible-sounding fictions, and that reviewer time comes directly out of time that would otherwise go toward real vulnerabilities in projects like Go, Angular, Bazel, and Fuchsia. Google's own public framing matched that math rather than panic - the announcement read as a routine administrative notice rather than a crisis statement - but outside commentary and press pickup read it more dramatically, framing it as evidence that AI slop is now breaking bug bounty programs outright, and drawing the most debate of any single post on the story.

Google is acting last among major programs forced into the same corner, not first. curl's maintainer ended a six-year-old HackerOne bounty on January 31, 2026 after estimating that roughly 20% of its submissions were AI slop, while only about 5% of the year's reports were genuine vulnerabilities - including a stretch where seven invalid reports landed within sixteen hours ^[3]. Two months later, HackerOne's own Internet Bug Bounty, which had funded open-source fixes since 2013, stopped taking new submissions entirely ^[3]. Google's OSS VRP, covering Go, Angular, Bazel, Protocol Buffers, Fuchsia, and critical third-party dependencies ^[2], is simply the largest and most visible name yet to hit the same wall, which suggests the triage economics are breaking industry-wide rather than reflecting a Google-specific management failure. That framing is already bleeding into mainstream tech commentary, where 'is bug bounty dead' has become its own recurring question independent of any single company's incident.

The reason these programs keep getting flooded is a cost asymmetry that generative AI makes worse every month: producing a bug report now costs the submitter almost nothing, while validating one still requires a human engineer's time. The money at stake explains why Google chose a pause rather than letting the noise pile up indefinitely - its VRP has paid out more than $81.6 million since 2010 ^[2], and 2025 alone saw $17.1 million distributed to over 700 researchers, a 40% jump from 2024 ^[4]. Individual OSS VRP rewards range from $100 to $31,337, with a separate Patch Rewards track paying up to $15,000 for verified fixes ^[2]. That is a program worth protecting rather than abandoning outright, which is likely why Google froze only the product-vulnerability intake instead of shutting OSS VRP down the way curl walked away from HackerOne altogether.

Google drew a narrower line than it might appear at first glance. The freeze applies only to new OSS VRP product vulnerability submissions - supply-chain reports and anything already in the queue keep moving ^[1]- and Google is actively steering researchers toward its Patch Rewards Program and Cloud VRP in the meantime ^[5]. The company has committed to an update in Q1 2027 ^[2], which leaves the real question unresolved. Security researchers discussing the pause online are already betting on an answer: that bounty programs respond by narrowing scope and gating access by reputation rather than staying fully open, since report volume, not report quality, is what now breaks a triage team's budget. Whether Google's Q1 2027 update confirms or avoids that path is the thing worth watching.