Generative AI was supposed to help security researchers find vulnerabilities faster. For Google’s Open Source Security Team, it has also created a less welcome outcome: an explosion of low-quality vulnerability reports that humans still have to investigate.
Google has paused new submissions to its Open Source Software Vulnerability Rewards Program (OSS VRP) while it redesigns the initiative. The decision follows a surge in AI-generated bug reports targeting open-source projects, many of which contain plausible-looking but incorrect vulnerability claims.
The problem has become common enough to earn its own name in the open-source community:AI slop.
AI Makes Reporting Bugs Almost Free
Google launched the OSS VRP in 2022 to reward researchers who discovered vulnerabilities in important open-source projects. The program eventually expanded to cover software such as Bazel, Angular, Chromium, Golang and Google Cloud-related repositories, with rewards reaching tens of thousands of dollars for serious findings.
The economics of vulnerability reporting have changed considerably since then.
Large language models can now inspect source code, generate security hypotheses and produce polished vulnerability reports with little human effort. A researcher once needed to understand a codebase, validate a vulnerability and write a convincing report. Today, an AI system can generate dozens of apparently credible findings in a fraction of the time.
Unfortunately, generating the report is much cheaper than verifying whether it is actually correct.
Maintainers receiving these submissions must still reproduce the claimed behavior, inspect the affected code and determine whether the issue represents a genuine security vulnerability. A hallucinated bug can therefore consume real engineering time even when the person submitting it invested almost none.
Open Source Maintainers Are Paying the Cost
Google is not the first organization to encounter the problem.
Curl project founder Daniel Stenberg has repeatedly warned about AI-generated security reports overwhelming open-source maintainers. Curl has received vulnerability submissions that appear convincing at first glance but collapse once developers investigate the underlying claims.
The Python Package Index has faced similar pressure. In August, PyPI temporarily stopped accepting new vulnerability reports after being flooded with low-quality submissions, many believed to have been generated using AI tools.
The asymmetry is becoming difficult to ignore:AI allows one person to generate reports at machine speed, while verification still happens largely at human speed.
That creates a denial-of-service problem for vulnerability disclosure itself. Even without malicious intent, enough inaccurate reports can consume the limited attention of maintainers and make genuine vulnerabilities harder to identify.
Google Is Redesigning the Incentives
Google says the OSS VRP pause is temporary and that existing submissions will continue to be processed. The company plans to introduce a redesigned program later this year, though it has not yet detailed exactly how the new model will work.
The challenge will be preserving the advantages AI brings to vulnerability research without rewarding automated speculation.
AI-assisted security research is already demonstrating genuine value. Specialized models can analyze large codebases, identify subtle relationships between weaknesses and help researchers develop exploit hypotheses that might otherwise require significant manual work.
But there is an important difference between using AI to assist research and using AI to manufacture vulnerability reports.
A credible submission still needs evidence: a reproducible attack path, clear security impact and enough technical validation to demonstrate that the vulnerability actually exists.
AI Is Changing the Economics of Security Research
Bug bounty programs were designed around an assumption that submitting a vulnerability required meaningful effort. Generative AI is breaking that assumption.
When the cost of producing a report approaches zero, bounty programs and disclosure platforms need stronger mechanisms for filtering quality. That could mean requiring reproducible proof-of-concept code, limiting submissions from researchers with poor validation histories or introducing automated triage before reports reach maintainers.
The irony is that AI may ultimately become part of the solution as well. Security teams could use models to evaluate reports generated by other models, creating another layer of automated filtering before human researchers become involved.
Google’s decision is therefore about more than one bug bounty program. It exposes a broader challenge emerging across cybersecurity: AI dramatically increases the amount of security information that can be produced, but it does not guarantee that information is useful.
For open-source maintainers already working with limited time and resources,the next security problem may not be finding more vulnerabilities — it may be separating the real ones from an endless stream of convincing AI-generated noise.