AI Bug Bounty Trust Collapse: The Real Reason Google Froze OSS VRP and 3 Structural Solutions

·

Key Summary

  • As of October 1, 2026, Google temporarily suspended its Open Source Software Vulnerability Reward Program (OSS VRP) and announced via an official post and X that an update would be provided in Q1 2027.
  • The company explicitly stated the reason as: “automated submissions, the vast majority of which are not valid.”
  • TechCrunch reported that cybersecurity experts had been warning for at least a year that AI slop could seriously damage bug bounty programs, and that this move represents the realization of those concerns.

The essence of the problem: as bug bounties morph into “automatable security labor,” the trust supply mechanism of reward programs itself teeters on collapse, and both program operators and maintainers are forced to bear the primary cost of filtering out AI slop.

Table of Contents

On October 1, 2026, Google announced through its official bulletin board and X that it would temporarily suspend new submissions to its Open Source Software Vulnerability Reward Program (OSS VRP). A timeline for returning with an updated policy in Q1 2027 was also disclosed. The company’s stated reason was a single line: “automated submissions, the vast majority of which are not valid.” Automated submissions have surged, and the vast majority of them are not valid.

The implications of that single line go far beyond a simple operational efficiency issue. It is a signal that the new trust cost created by the buzzword AI bug bounty has begun to shake the security ecosystem itself.

Timeline: The One-Year-Old Warning Becomes Reality

According to TechCrunch’s report, cybersecurity researchers had been warning for at least a year that LLM-generated fake reports could overwhelm bug bounty triage pipelines. As early as mid-2025, comments such as “Report quality has dropped noticeably since GPT-5 release” had already appeared dozens of times in HackerOne and Bugcrowd community threads. Yet the broader industry dismissed this as natural noise in the cybersecurity market, and no operator actually shut down an entire program over it.

Google’s OSS VRP became the first casualty. Immediately after the early October announcement, testimonies began pouring in from major open source maintainers on GitHub: “We’ve been suffering from the same symptoms for days now.” Invalid reports flooded inboxes and issue trackers, and the cost was passed on not to program operators but to the project maintainer community. According to the original TechCrunch article, Google’s suggested workaround is switching to its other VRPs (Google VRP, Abuse VRP). But open source projects don’t have that workaround available.

3 Structural Damages Caused by AI Slop

The first is exploding triage costs. Based on the author’s hands-on experience running vulnerability triage, processing one valid report takes an average of 30 to 90 minutes. Yet LLM-generated fake reports often take even longer than that. That’s because PoC reproduction is impossible, affected versions are filled in by guesswork, and non-existent functions are cited. A triager must fact-check the entire body before finally reaching an “invalid” conclusion.

The second is maintainer fatigue. Most core OSS project maintainers perform maintenance part-time while holding other jobs. When dozens of invalid reports pile up per day, they end up spending more time on classification than on code contribution. Over the long term, maintainers themselves drop out, and as a result, the open source security safety net weakens. The author has observed one Node.js library maintainer publicly confessing recently: “I’m closing incoming new submissions in the issue tracker.”

The third is dilution of the genuine reporter signal. When there were 10 reports, finding the 1 real one among them took 10 minutes. When reports jump to 100, finding that same 1 real one takes 2 hours. The cost doesn’t grow linearly; it grows exponentially. This loss is borne by program operators, real researchers, and end users alike.

Why Existing Approaches Fail

HackerOne and Bugcrowd introduced embedding-based duplicate detection and reporter reputation systems starting in 2024, but these alone were insufficient. LLM-generated reports are superficially very similar to legitimate reports. “Version notation, CWE category, and impact description” are filled in plausibly, so they are likely to be classified as “new issues” under simple embedding similarity checks. Moreover, because LLMs produce different expressions on every call, it’s easy to automatically generate patterns that resubmit the same flaw in varied forms.

More critically, operators have failed to reach consensus on how to define a report as “AI-written.” If an LLM claiming 90% accuracy merely authored some sections, is that a synthetic report or an assisted report? The very ambiguity of this boundary makes first-stage filtering difficult.

A 3-Stage Quality Gate Ready to Deploy in Practice

There is no perfect solution, but practical workflows that structurally reduce triage burden already exist. The key is enforcing first-stage evidence at the submission phase.

First, reproducible PoCs and HTTP transaction captures should be made mandatory fields. Require attachments of raw request/response logs captured by tools like Burp Suite or mitmproxy, along with execution commands such as docker run --rm -it vulnerable-app:latest and their output. LLMs struggle to consistently produce first-stage evidence at this level. Using Semgrep to statically inspect the reported code location and verify whether the actual function call path exists is also effective enough as a first-stage gate.

Second, CVE-ID and CWE mappings should be automatically cross-verified. Call the CVE 5.0 API at cve.org or the NIST NVD cves?cveId= endpoint to rule-based verify whether the CVE listed in the report actually exists, and whether an item classified as CWE-79 has been incorrectly labeled as “RCE.” Comparing Trivy’s --format json output against the reported affected library versions falls in the same category. When score-evidence inconsistencies are detected, route them to the automated human triage queue, and automatically downgrade priority for new or low-reputation reporters.

Third, reputation-weighted queues should be formalized. Combine a reporter’s past submission accuracy, operator response rate, and false positive ratio into a reputation score, and delay new submissions from accounts with scores below a certain threshold by 24 to 72 hours. HackerOne has already partially introduced this mechanism, but open source projects implementing it on their own can build a fully sufficient workflow by combining GitHub’s .github/SECURITY.md with GitHub Actions. The key is making trust-weighted signals, not raw submission volume, the primary input for triage.

How Will Google’s AI Bug Bounty Policy Change in Q1 2027?

It is difficult to assert anything definitively based only on publicly available information, but the direction can be estimated. First, PoC-based submission will very likely become the de facto standard. Second, reporter identity verification and prior reputation checks will likely be introduced. Third, policies explicitly requiring declaration of AI-assisted tool use, i.e., self-reporting items such as “which parts of this report were written by an LLM,” may be included. OpenAI and Anthropic have already published guidelines recommending disclosure of AI-assisted work.

Industry standardization is also an open issue. Follow-up amendments to ISO/IEC 29147 and 30111 are likely to address LLM-generated reports, and related items are expected to be added to the OWASP Bug Logging Guide. In Korea, KISA’s vulnerability disclosure reward program operating guidelines may face similar pressure going forward, so domestic security officers need to reflect PoC capture and CVE cross-verification routines in their internal SOPs in advance.

What to Watch For

  • PoC enforcement level in Google’s new Q1 2027 policy — Whether they require only basic HTTP capture or full containerized reproduction environments will determine the industry-wide baseline.
  • Whether HackerOne and Bugcrowd disclose their reputation systems — If they publish the weighted score calculation transparently, smaller operators can adopt similar models.
  • Standardization of open source maintainer protections — Whether submission caps and invalid-report accumulation policies are provided at the platform level by GitHub, GitLab, etc.
  • Impact on domestic KISA and KHTML vulnerability reporting systems — When and how global trends are reflected in Korea’s system design.

Closing

The moment bug bounties morph into “automatable security labor,” the trust supply mechanism of the program itself stops working. LLMs can produce reports quickly, but they do not bear responsibility for the actual risk of the vulnerabilities those reports point to. That responsibility ultimately falls on operators and maintainers. The policy Google brings forward in Q1 2027 should be not just a simple gate addition, but a redesign of trust signals. The security ecosystem must, once again amid the noise of AI bug bounties, return to the essential question: “Whose evidence, and from whom, do we trust?”

Frequently Asked Questions

I’m an open source maintainer. What should I do right now?

First, clearly state in your SECURITY.md that a reproducible PoC and HTTP transaction capture are mandatory requirements. At the same time, leverage GitHub’s repository security advisories feature and issue templates to deploy a GitHub Action that auto-labels invalid reports — this can significantly reduce classification workload.

Won’t sanctions on new or low-reputation reporters penalize legitimate researchers?

Design it not as a simple block, but as a 24 to 72 hour delay plus additional verification requirements. New submitters who provide valid PoCs join the normal queue after the delay, which is more equitable than identity-based blocking. The point is not “we want fewer submissions,” but “we want only submissions backed by trust signals.”

How is legitimate security research using AI tools protected?

Most policy drafts do not prohibit LLM use itself; they require self-declaration of use. In other words, using AI as an assistive tool is normal, but submitting AI-generated false claims as-is is the behavior subject to sanction. Once standards are established, it is highly likely that report templates will add an “LLM use status and scope” field.

Source

This article was prepared after reviewing the following original: TechCrunch — Google froze its open source bug bounty program due to a ‘significant rise’ in AI submissions

Leave a Reply

Your email address will not be published. Required fields are marked *