Google has already used AI-assisted fuzzing to find 26 vulnerabilities, including one in OpenSSL. Now it has paused part of its open-source bug bounty because automated reports are flooding the inbox with invalid claims.
Same technology. Very different outcomes.
The useful dividing line isn’t human versus machine. It’s whether somebody tested the thing before asking a security engineer to investigate it.
The inbox closed, not the whole program
Effective October 1, 2026, Google stopped accepting new product-vulnerability submissions to its Open Source Software Vulnerability Rewards Program, or OSS VRP. As [Tom’s Hardware reports](https://www.tomshardware.com/tech-industry/artificial-intelligence/google-suspends-part-of-the-oss-vrp-bug-bounty-program-due-to-an-influx-of-invalid-ai-submissions-product-vulnerability-submissions-ended-october-1), earlier submissions are unaffected, and supply-chain reports remain eligible. Some vulnerabilities affecting Google Cloud products can still qualify through Cloud VRP.
Google’s explanation is blunt:
<> “This pause is due to a significant rise in automated submissions, the vast majority of which are not valid.”/>
The company promises an update in Q1 2027. That’s an update, not a reopening date.
Calling this a shutdown of Google’s bug bounties would be wrong. Across its rewards programs, Google paid more than $17 million to over 700 researchers in 2025, up more than 40% from 2024. This is a blocked intake lane inside a substantial security operation.
Still, blocking that lane is a serious admission: the submission process has become expensive enough to interrupt legitimate research too.
A convincing report isn’t a working exploit
Google had already tried raising the bar. Its [March 2026 OSS VRP rules](https://bughunters.google.com/blog/ossvrp-rule-updates-2026) required exact reproduction through an existing OSS-Fuzz target or a merged patch for memory-corruption reports in flagship and important projects. April brought narrower reward eligibility for lower-priority projects.
Then came the pause.
Google describes two recurring problems: hallucinated trigger conditions, and real coding errors without meaningful security impact. An overflow in unreachable code can look alarming in a beautifully formatted report. It still doesn’t give an attacker a reachable attack path.
This is where polished prose becomes dangerous. A report can include CWE labels, ominous severity language, and a tidy remediation section while skipping the only interesting question: can an attacker make this happen?
For developers using AI to investigate code, the useful output is evidence:
- A minimal reproducer against the affected version.
- Exact commands, environment details, and observed results.
- An explanation of attacker-controlled input and the security boundary crossed.
Everything else is supporting material. Your chatbot’s confidence is not a stack trace.
What Nobody Is Talking About
Better reports won’t necessarily fix the workload problem.
Daniel Stenberg’s experience with curl makes that uncomfortable point. He ended curl’s bounty on January 31, 2026, after it had produced 87 confirmed vulnerabilities and paid more than $100,000 since 2019. The confirmation rate had fallen from above 15% in earlier years to below 5% in 2025.
But in April, Stenberg described a different situation: the slop had subsided, while report frequency was higher than ever. His name for it was “High-Quality Chaos.”
That’s the harder problem. Even valid findings need reproduction, prioritization, fixes, regression tests, and release coordination. Generating discoveries faster does not magically generate maintainer hours.
Google’s pause is defensible as a pressure-release valve. It’s also a blunt instrument: genuine researchers lose a reward route alongside people submitting speculative nonsense.
The better investment is verification infrastructure, not another agent that produces suspicious-code essays. Google’s own [AI-fuzzing results](https://security.googleblog.com/2024/11/leveling-up-fuzzing-finding-more.html?m=1) show the difference: generated hypotheses became findings through executable tests.
AI can help find bugs. But if the submitter skips validation, the automation hasn’t eliminated work. It has handed that work to a maintainer who never agreed to do it.
