Key Summary
- Google has frozen the Open Source Software Vulnerability Rewards Program (OSS-VRP) effective October 1, 2026, and announced that an update will be provided in Q1 2027.
- The official reason—”a sharp rise in automated submissions, the vast majority of which are invalid”—was posted simultaneously on X (formerly Twitter) and the program’s official website.
- Internally, Google engineers and open source maintainers were overwhelmed by AI-generated invalid reports and submissions containing hallucinations, according to Tom’s Hardware.
The credibility of a bug bounty program begins not with “who reported it” but with “is the report reproducible.” The more AI scales up submissions, the more automation of the triage pipeline that demands verifiable evidence becomes a core competitive advantage.
Table of Contents
- Key Summary
- Background: How AI Noise Neutralized Existing Triage
- Practical Response 1 — Input Gate That Enforces PoC Reproducibility
- Practical Response 2 — Isolating Low-Quality Submissions with Reputation Scores and Cooldowns
- Practical Response 3 — LLM-Assisted Triage, Human Gate Preserved
- Connection to the Governance Restructuring Trend
On October 1, 2026, Google simultaneously announced on X and its official website that it was temporarily freezing the Open Source Software Vulnerability Rewards Program (OSS-VRP). New submissions will not be accepted until the Q1 2027 update, with the official reason given in a single line: “a sharp rise in automated submissions, the vast majority of which are invalid.” According to TechCrunch’s report, internally, engineers and maintainers were overwhelmed by LLM hallucination reports and duplicate automated scanner submissions. From the perspective of a bug bounty triage operator, this incident is not a simple temporary halt—it reveals a structural defect in the reward system itself, which depends on external signals.
Background: How AI Noise Neutralized Existing Triage
The “AI slop” that cybersecurity experts warned about in 2025 has become reality within a year. The flow of LLMs using grep to scrape patterns from arbitrary repositories and crafting plausible exploit scenarios for submission has grown. A single automated scanner throwing out hundreds of reports per day burns dozens of hours per week of human first-line filtering. Not only Google engineers but also individual maintainers’ personal notification channels have been flooded with invalid reports, creating the side effect of burying genuine vulnerability signals.
The biggest problem is plausibility. Existing triagers look at: ① the logical coherence of reproduction steps, ② the dependency references in code snippets, and ③ the accuracy of the affected version notation. LLM reports generate plausible sentences in ① and ③, neutralizing the first-line gate that relied on the intuition of experienced engineers. Teams that rely on humans to handle the first-line filter without an automated parser eventually become paralyzed. In my experience with similar patterns in the field, when the triage queue is flooded with invalid reports, top reporters voluntarily leave, rapidly eroding ecosystem trust.
Practical Response 1 — Input Gate That Enforces PoC Reproducibility
The first place to start is at the report intake stage. Enforce a template via JSON Schema and make fields like environment.image_digest, steps[].command, and logs.snippet required. In GitHub Actions, after running pip install jsonschema, automatically validate with jsonschema.validate(instance=payload, schema=schema), and reject submissions that don’t meet the format, telling the submitter exactly which fields are missing.
For reproduction step validation, use CodeQL and Semgrep. If the attached snippet violates semgrep --config p/security-audit rules, or if the symbol isn’t actually imported in the CodeQL database, flag it as “non-reproducible.” In my experience, 60–70% of the noise is automatically filtered out at this stage. The key is to fully automate everything that can be mechanically verified before consuming human time. Once the validation logic is done, only the results flow into the human queue.
Practical Response 2 — Isolating Low-Quality Submissions with Reputation Scores and Cooldowns
Enforcing reproduction alone is not enough. The pattern where the same author repeatedly submits meaningless scenarios that merely satisfy the format must be handled by a separate system. In Redis, store reporter:{id}:score and reporter:{id}:last_submission_at, and compute a reputation score between 0 and 100 by combining weights for acceptance rate, reproduction success rate, and domain expertise. Submissions below a certain score are forcibly moved to the manual review queue, and repeat low-quality submitters face a progressive cooldown of 24 hours → 72 hours → 7 days. This logic fits in a few dozen lines but has significant operational impact.
In environments running container image scans with Trivy, require trivy image --severity HIGH,CRITICAL results to be attached as part of the PoC, naturally weeding out simple LLM hallucinations. Teams without a container environment can achieve the same effect by including dependency CVE matching via OWASP Dependency-Check in the PoC. Both tools are free and open source, so the adoption friction is minimal.
Practical Response 3 — LLM-Assisted Triage, Human Gate Preserved
Automatic acceptance is absolutely prohibited. LLMs only handle first-pass classification and clustering, and humans make the final determination. Embed reports with sentence-transformers (all-MiniLM-L6-v2), then auto-cluster with existing reports that have a cosine similarity of 0.85 or higher, sending only the first report from each cluster to the human review queue. For code patterns, extract the diff with git log -p and analyze it with CodeQL to confirm whether the vulnerable function signatures match. The goal at this stage is to reduce the number of items entering the human queue by 70% or more.
At the human review stage, actually execute the PoC in a sandbox container isolated by k8s NetworkPolicy. After default-deny on policyTypes: Ingress,Egress, whitelist only the images that are explicitly needed, and sandbox the runtime with gVisor or firecracker. If the results are automatically attached to the report, reviewers can make a judgment within 5 minutes. The most commonly overlooked pitfall in practice is the model’s own hallucinations. Trust the classification result, but leave the rationale for the final decision in human-written comments, so that model bias can be tracked when analyzing false positives later.
If the clustering threshold is set too low, real new vulnerabilities can be bundled and disappear. It’s safe to start conservatively at 0.9 or higher for the first two weeks, and adjust by monitoring the false positive rate on a quarterly basis. KPIs that can be made public are the auto-block rate, average triage time, and false positive rate, and even when operated privately, quarterly summaries should be disclosed to participants to maintain trust.
Connection to the Governance Restructuring Trend
This freeze is not an isolated incident. Just as agent governance cases have exposed the limits of external arbitration, bug bounties are also structurally vulnerable to model-generated noise as long as they rely on signals from external participants. As AI penetrates deeper into policy and security domains, the same pattern will repeat in other programs unless the community itself has its own control mechanisms. Until Google’s Q1 2027 update, it’s time for field teams to focus on modernizing bug bounty triage.
Issue Summary
- Input gate: Enforce report format with JSON Schema and use CodeQL and Semgrep for first-line blocking of non-reproducible reports. The key is to fully automate mechanical verification before human review.
- Reputation & cooldown: Strengthen PoC substance with a Redis-based reputation score, IP/account-level cooldown, and Trivy/OWASP Dependency-Check. The balance of not blocking new reporter entry while isolating only low-quality repeats is important.
- LLM-assisted triage: Combine embedding clustering with sandbox dynamic validation, but prohibit auto-acceptance and always maintain the human review gate. Start with a conservative threshold and adjust the false positive rate on a quarterly basis.
- Governance: AI-generated noise is not a problem for a single program, but a structural challenge for the entire reward system that relies on external signals. Control mechanisms must be established while maintaining trust with public KPIs.
Frequently Asked Questions
Q. Where should small teams with limited budgets for automation tools start?
Enforcing a JSON Schema-based report template (one engineer can build it in a day), automatic code snippet validation with Semgrep’s free rule set, and requiring Trivy PoC attachments—these three alone filter out a significant portion of the noise at the first stage. If there’s no container environment, including dependency CVE matching via OWASP Dependency-Check in the PoC is the next best option.
Q. How do we ensure the reputation system doesn’t penalize existing top reporters?
Design it so that new reporters can rise sufficiently with 2–3 verified submissions. The key is not “starting low” but “demoted when low-quality repeats accumulate.” Demotion thresholds and recovery conditions should be specified in advance to increase predictability.
Q. How do we securely configure sandbox dynamic validation?
In a separate namespace, default-deny all Ingress/Egress with NetworkPolicy and whitelist only the necessary images. Isolate the CRI runtime with gVisor or Kata Containers, and mount host directories as read-only. Preserve PoC execution logs for audit trails.
Source Article
This article was prepared after verifying the following original source: TechCrunch — Google froze its open source bug bounty program due to a ‘significant rise’ in AI submissions

Leave a Reply