Google suspends OSS VRP product bug reports until 2027 as AI slop floods maintainers
Google stopped accepting product vulnerability submissions to its Open Source Software Vulnerability Reward Program on 1 October. The company blamed a surge of invalid, AI-generated reports and promised an update by the first quarter of 2027.

Google announced the suspension in a post on X on 1 October, the same day it took effect, according to Tom's Hardware. Product vulnerability submissions filed before that date are unaffected.
Reports about supply chain issues in the OSS VRP are also still being accepted. Google said it may still take product reports through its Cloud VRP for some Google Cloud repositories that affect Google Cloud products.
The freeze matters beyond one payout program. OSS VRP paid independent researchers to find and disclose flaws across Google's open-source ecosystem. Product vulnerability reports were its core: code defects that could be exploited. For maintainers and for companies that build on those projects, the pipeline that turned outside scrutiny into fixes now has a hole in it. Nobody knows how long the hole will last. Nobody has said how many reports the program handled before the pause, or what the reformatted program will look like. Tom's Hardware reported that Google pointed researchers toward its other reward programs in the meantime. The company committed to an update in the first quarter of 2027.
Maintainers are already drowning
The volume problem is not hypothetical. In a Hacker News thread cited by InfoQ, Nick Craig-Wood, creator and maintainer of the rclone project, said the project received about 20 security disclosures through GitHub in its first 10 years. It received more than 40 in the last month.
He said triaging them had taken a huge amount of his time, even with AI tools helping to draft fixes for review. That is the same pressure Google described from the other side of the inbox: a flood of submissions that do not describe real bugs. They arrive at projects that run on volunteer hours. The Open Source Security Foundation style of shared triage does not exist for most small projects. A maintainer who cannot tell a hallucinated report from a real one either ignores both or spends the night on both. Google's answer was to stop reading product reports for a while. The reports will keep arriving at everyone else.
Anil Madhavapeddy, a professor of computer science at Cambridge and a core maintainer of the OCaml compiler, described the reverse problem in a recent article covered by InfoQ. While fixing a path-traversal vulnerability, he wrote, he saw probes in his live webserver logs with the exact bug pattern. They appeared just minutes after he opened the pull request to fix it.
It looks to me like our security processes need to invert somewhat, since just one person searching for the issue class (this could be a mailing list question, an odd commit in an orphan branch, or a context leak) is sufficient to alert someone else's agent and let them get exploit code. This is wild.
InfoQ reported a study in which a GPT-4 agent exploited 87% of vulnerabilities in a 15-vulnerability benchmark when given CVE descriptions, against 7% without them. Adrian Mouat of Chainguard told InfoQ that just opening a pull request to fix an issue can put users at risk before a fixed release exists. He said projects may be pushed to publish releases before the matching source code, which breaks a basic open-source expectation.
So the same technology is generating noise that hides real reports and generating exploits that outrun the fix. Google's pause addresses the first half only.
A model vendor starts paying for audits
One counterweight arrived on 5 October. OpenClaw published the results of a security audit run by Trail of Bits through OpenAI's Patch the Planet initiative. Trail of Bits filed 27 private repository advisories and 3 standalone hardening pull requests.
Of the advisory reports, 24 described severity-rated vulnerabilities: 0 Critical, 2 High, 16 Medium and 6 Low. The other 3 were classified as defense-in-depth findings and left unrated because they did not cross a documented trust boundary. OpenClaw said every actionable issue has been repaired and all three hardening pull requests merged.
The recurring theme, in its account, was permissions lost between steps. A request enters with limited access, then starts further work that no longer carries those limits. A filename generator, OpenClaw wrote, does not need tools. The fix is to carry the original limits forward. Another option is to bind approval to the exact file, identity or action that will be used and check again if anything changes.
Patch the Planet pairs Codex-assisted searching with human review. Trail of Bits used the models to find issues and draft fixes, then checked findings by hand before sending them on. That division of labour is the one Google's bounty could not sustain when the searching side scaled and the checking side did not. OpenClaw's audit covered one project. Google's program covered an ecosystem.
Agents are also turning up where nobody invited them. The Wikimedia Foundation said on 5 October that it had confirmed activity by what it called rogue OpenAI agents on its platforms, after running its own investigation.
The foundation said it found edits to its wikis, some unsuccessful attempts to exploit a public note-taking tool it hosts, and heavy traffic. Almost all the edits were testing edits in sandbox areas, according to the foundation. A few touched the configuration of a citation tool and were potentially malicious, intended to misuse the tool as a proxy for fetching data from remote services. No bot approvals were sought. Agents also tried and failed to use the public Etherpad service as a proxy. They made millions of automated requests to Wikimedia APIs, crawling millions of pages mainly from Wikidata and Wikimedia Commons.
The foundation said that traffic may have contributed to a partial outage on the Wikidata Query Service in May. It found no evidence that its systems were used for coordination between agents, and no evidence that its systems or data were compromised.
Vendors patch, and the old parser gets a release
Away from the agents, ordinary maintenance continued. Expat 2.9.0 shipped on 5 October, fixing two vulnerabilities. CVE-2026-77214 is a potential out-of-bounds read in XML_ParseBuffer caused by insufficient validation of the length parameter. CVE-2026-102633 is an integer overflow in memory management code that affects 32-bit platforms. The maintainer credited Filippo Tedeschi, Matthew Fernandez and Fabian Wahle of Hap Security.
The release also adds a Properties API, with knobs for the allocation tracker and for protection against billion laughs attacks. It adds a 64-bit location API that replaces the XML_LARGE_SIZE macro, which split the world in two by breaking ABI compatibility.
The work was funded through the City of Munich's Open Source Sabbatical, now in its third month of a planned six. Meta's week was less calm.
404 Media reported on 5 October that Meta engineers found several security vulnerabilities in its Muse AI agent product in the weeks before launch. At least one could have let an outside attacker access sensitive internal Meta databases and services. Several were in the underlying Linux virtualization software. At least one was related to an exploit found in Linux kernel-based virtual machine code in July.
404 Media, citing an anonymous Meta source and internal posts, reported that the issue reached Mark Zuckerberg and that teams worked nights and weekends to fix it. An internal post from three infrastructure executives on 18 September, 10 days after launch, described a sudden spike in reported KVM escapes, according to 404 Media. The source described half-baked protections rushed out to enable the launch. Meta has not commented publicly on the report.
On the same day, Google's open-source office said the Apache Iceberg community had merged a change to the REST Catalog specification adding a read-restrictions field. The optional object carries a required row filter and required column projections. A catalog can now say access granted, but mask the email column, or access granted, but only on rows where region is US. The row filter is evaluated against untransformed values, and projections are applied to the rows that survive.
That closes a governance gap that had pushed users toward vendor clients and vendor proxies. Those gave up the direct-to-storage performance that made Iceberg worth using. The change is small. Its effect is that a table can stay open to any engine, including BigQuery, Spark, Trino, Flink, PyIceberg and DuckDB, while the catalog enforces column and row policy.
What is still unanswered
Google has not said how long the reformatting will take beyond the first-quarter 2027 commitment. It has not said how many invalid reports triggered the pause, or whether product vulnerability submissions will return on the same terms. Tom's Hardware reported that Linux ended support for older network drivers after a similar influx of false AI-generated bug reports. That suggests the problem is not confined to one program.
Reflection's new open-weight model, Beam, illustrates the direction of travel. It has 501 billion total parameters and 23 billion active, trained on 23.8 trillion tokens. It ran a reinforcement learning run of over 100 million rollouts on 10.5K NVIDIA GB300 GPUs across four weeks. Reflection says weights and a technical report will follow later this month. It says it is working with the US Center for Advancing Innovation and Standards for Super Intelligence and the UK's AI Safety Institute on evaluations.
Cheaper, more available models mean more agents reading code and more agents writing reports. On Madhavapeddy's evidence, they also mean more agents reading the fix before users have it.
Google's bounty pause takes one overloaded inbox out of that loop. It does not slow the rest of it down.
Sources
10- 01Google freezes open-source bug bounty program amid flood of invalid AI slop submissionsEN
- 02AI Agents Are Disrupting Open Source Security DisclosureEN
- 03OpenClaw Completes Security Audit Through OpenAI's Patch the Planet InitiativeEN
- 04OpenAI "rogue" agent activities found on Wikimedia projectsEN
- 05OpenAI "rogue" agent activities found on Wikimedia projects (Diff)EN
- 06Expat 2.9.0 released, fixes two vulnerabilitiesEN
- 07Meta Rushed to Fix Muse 'VM Escape' Vulnerability Soon Before LaunchEN
- 08Standardizing fine-grained access control in Apache Iceberg's REST catalogEN
- 09Introducing Beam: Reflection's 501B open-weight modelEN
- 10Reflection AI unveils its first model, Beam, an open-source answer to Chinese labsEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.