Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Open source security is drowning in AI slop, and bug bounties are paying the price

Google stopped accepting new product vulnerability reports to its Open Source Software Vulnerability Reward Program on 1 October. It cited a flood of invalid AI-generated submissions. The pause runs until at least the first quarter of 2027.

TechnologyAnalysisRachel NwosuPublished: 4 October 20265 min readSources 8
Open source security is drowning in AI slop, and bug bounties are paying the price

Google announced the suspension in a post on X on 1 October, Tom's Hardware reported. The change took effect the same day. It does not touch reports filed before that date.

Google said it may still take product vulnerability reports through its Cloud VRP for some Google Cloud repositories. Supply chain reports under the OSS VRP remain open. That matters to anyone who runs open source code, which is most of the internet. Maintainers get less free scrutiny from outside researchers, at least for now. Google has promised an update by the first quarter of 2027. It has not said what a rebuilt program will look like, or how much of the old coverage it will restore.

The program pays independent researchers to find and disclose flaws across Google's open source ecosystem. Product vulnerability submissions cover code defects. Supply chain reports cover a different set of problems in the dependencies that projects pull in. Google's own wording, as reported by Tom's Hardware, is that it is reformatting this part of the program.

Why the reports stopped being useful

The trigger is volume. Tom's Hardware reported that the program was suspended over an influx of invalid AI-driven reports. It is not an isolated complaint. The same report noted that Linux dropped support for older network drivers after a wave of false AI-generated bug reports. That is the same failure mode at a smaller scale.

Two reports in the past few days sketch the two sides of this. On 2 November, a public tracker of CVEs credited to Anthropic's research team listed Apache Arrow in Rust with a CVSS score of 9.9. It also listed three flaws in NLnet Labs Krill rated 8.4. Those entries sit under the same tracker that records SQLite at 6.8 on 30 October and Nextcloud Server at 5.3 on 27 October, according to the GitHub repository maintained by patrickmgarrity.

So AI-assisted review does find real, serious bugs. The problem is the ratio. A bounty program has to triage every submission. A report that turns out to be a hallucination costs the same reviewer time as a real one.

That asymmetry is the whole story. Finding candidates got cheap. Deciding which ones are real did not.

Researchers are building the filters themselves

Caleb Gross, an offensive security researcher, published a talk transcript on 4 November. It describes tools built for exactly this triage problem. His Slice tool reproduced discovery of a use-after-free in the Linux kernel SMB server for about $3 per run, according to his writeup at noperator.dev. His Raink tool ranks large data sets with LLMs in O(N) complexity. He says it can prioritise roughly 2,700 GitHub repos, about 3,000 kernel subsystems or around 1,500 patch diff functions.

The framing in Gross's talk is prioritisation, not detection. He argues the bottleneck is deciding which attack surface to examine, not the bug-finding itself.

Industry practice is starting to agree. Google's freeze is the blunt version of that judgment: if the queue cannot be ranked, close the queue.

What the closed queue hides

One vendor's pause is not the same as a fall in vulnerability discovery. The Anthropic-linked tracker shows a steady stream through late October and early November. The Register reported on 3 October that Anthropic's bug-hunting model Mythos was involved in a vulnerability that has since come under attack. That suggests the tooling keeps producing findings whether or not Google's program is open.

The risk sits with projects that relied on bounty attention as their only external review. A small library with no security team and no paid audit gets nothing from a program that is closed, even if the closure is temporary.

Google has not said how many reports the program handled before the pause, or what share were invalid. Neither has it said whether the program's budget changes. Those numbers would tell maintainers how much scrutiny they are actually losing.

The Cloud VRP carve-out is narrow. It covers some Google Cloud repos whose flaws affect Google Cloud products. It does not restore general coverage of open source code that Google does not operate as a service.

The parallel argument about how bugs get found

The same week brought a separate argument about who should be looking. David Robinson led the writing of safety reports that accompanied OpenAI product releases. He resigned and wrote in the Atlantic that the company's culture is broken, the Guardian reported on 3 October. He argued that AI firms are not being careful enough. He called for safety expertise from fields such as nuclear and aviation.

Robinson's essay is about model safety, not bounty programs. It should not be stretched to cover Google's decision. But both stories turn on the same question: what happens when the volume of machine-generated output outruns the humans asked to check it.

For open source maintainers, the practical answer this week is that one large source of free review has gone quiet. For researchers who filed real reports before 1 October, nothing changes. For everyone else, the queue is closed.

Google's commitment is an update by the first quarter of 2027. Until then, the OSS VRP supply chain channel stays open. The Cloud VRP covers a slice of Google Cloud repositories. Neither replaces the product vulnerability pipeline that closed on 1 October.

Comments 0

Sources

8
  1. 01Google freezes open-source bug bounty program amid flood of invalid AI slopEN
  2. 02Tracking vulnerabilities that credit the Anthropic research teamEN
  3. 03O(N) the Money: Scaling Vulnerability Research with LLMsEN
  4. 04OpenAI safety leader quits, warning AI company's culture is 'broken'EN
  5. 05Xray-core concealed a certificate verification bypass vulnerabilityEN
  6. 06Command injecting a reference tool to copy a source fileEN
  7. 07PhotoSuite Is an Open Source Photoshop-Like Editor with Native PSD SupportEN
  8. 08OpenBSD Developers Reject uutils Coreutils Port Over Licensing and Compatibility ConcernsEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Rachel Nwosu

Rachel Nwosu

AI, models and technology

Rachel Nwosu covers AI, models and technology for FLASH24, working from public model documentation, benchmark releases and repository histories rather than press summaries, and she skips announcements that arrive without reproducible numbers. She checks training-data claims against dataset cards and reruns reported metrics where code is available. She spends much of her week interviewing researchers and engineers, tracking model launch calendars, and comparing vendor benchmarks with independent evaluations. Outside the desk she runs 3D printers, restores old computers, and tests how models learn from internet junk. She does not publish benchmark figures she cannot trace to a source.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.