VulnHunter Can Prove a Flaw Is Real. It Can’t Prove Anyone Owns It

Capital One open sourced a tool that tries to disprove its own vulnerability findings before an analyst ever sees them. A Treasury-led federal clearinghouse now coordinates AI-powered vulnerability discovery across the entire government. Neither one answers who is accountable when the finding outruns the fix.

TL;DR

• Capital One open sourced VulnHunter, an AI tool that maps attack entry points and tries to disprove its own findings before alerting analysts.
• The US Treasury launched Gold Eagle, a federal clearinghouse coordinating AI-powered vulnerability discovery, validation, and patching support across agencies, companies, and open source maintainers.
• Both are real improvements in finding and validating vulnerabilities faster. Neither answers who is accountable for what happens after a finding is validated.
• CVE submissions are up 263% since 2020, with a widening gap in NVD's enrichment pipeline, alongside sharp increases in malicious open source packages tracked independently by ReversingLabs and Sonatype.
• The accountability decision, who owns a validated finding and on what documented timeline, is a governance decision, not a detection upgrade, and most organizations have not made it.

I have sat in the room where a security leader walks a board through the vulnerability backlog for the first time. The number is always large. The explanation is always the same: we know about these, we are working through them, the queue is the queue. Nobody in that room asks the question that actually matters, which is who decided the queue would be reviewed by people in the first place, and whether that decision still holds now that the queue can fill itself faster than any team can read it.

Capital One's new open source security tool, VulnHunter, and the US Treasury's newly announced clearinghouse, Gold Eagle, landed in the same stretch of the news cycle for what is really one reason. Both are betting that finding vulnerabilities faster, with AI doing the finding, is the problem most worth solving right now. VulnHunter starts from a likely attack entry point, maps the exploit chain an attacker would actually use, and tries to disprove its own finding before it reaches a human analyst. Gold Eagle does something structurally similar at government scale: it coordinates AI-powered vulnerability discovery across agencies, validates the findings, and connects the resulting queue to companies and open source maintainers who can act on it. Both are, in their own register, an answer to the same complaint security teams have carried for years, that most of what lands in the queue is noise and the real signal is buried somewhere underneath it.

That is genuine progress, and it is worth saying plainly. A tool that tries to disprove its own exploit chain before an analyst sees it is a better tool than one that does not. A federal body that can tell a small open source maintainer their package has a validated, exploitable flaw, instead of leaving them to learn it from a security researcher's blog post, is a real improvement in how that maintainer's week goes.

None of that answers the one question a board should actually be asking, which is who is accountable for what happens after a finding is validated. Gold Eagle can tell an agency or a company that a flaw is real. It cannot make that agency patch it on any particular schedule, and the early reporting on its launch already says as much: questions remain about oversight, how sensitive disclosures get handled, and whether validated findings turn into completed patches quickly enough. VulnHunter can hand Capital One's analysts a de-noised, validated finding. It cannot decide, on its own, which of those findings gets fixed this sprint, which gets filed for later, or who signs off on that tradeoff when a customer asks about it after the fact.

That gap is not new. It is the same gap that shows up every time an organization adopts a scanner, a static analysis tool, or a dependency checker: the tool changes what gets found, not who is responsible for what happens next. What is new is the speed. AI-generated code and AI-assisted vulnerability discovery are accelerating the rate at which validated findings arrive, on both sides of the fence, attacker and defender alike. The governance layer that was supposed to decide what happens to a finding once it exists was built for a world where a human security engineer generated a handful of validated findings a week. That world is gone. Nobody replaced the decision-making layer underneath it. They just made the layer above it faster.

The same shape shows up everywhere agentic AI is deployed right now, not only in vulnerability scanning. A locally run AI coding agent recently found to be capable of escaping its own sandbox and reaching host credentials was not a failure of intent. It was a tool built and shipped faster than the review process that was supposed to decide what access it actually needed. An autonomous email-connected agent quietly created through a single phishing link, inheriting a victim's existing business permissions, is the identical failure wearing different clothes. An agent that can act faster than the review process behind it is a governance question before it is anything else, and that question is not one any of these tools answer on their own.

I have watched this exact failure mode play out in the software supply chain for far longer than these announcements have existed. Enterprises built up open source dependencies one developer decision at a time, for a decade, without ever holding a board-level conversation about who was accountable for what got pulled in. Now those same organizations are watching AI accelerate that consumption, and the accountability question that was never resolved for the slow version of the problem is being asked, more urgently, about the fast version. NIST's own data shows CVE submissions up 263% between 2020 and 2025, alongside a gap in the National Vulnerability Database's enrichment pipeline that has only widened since April. ReversingLabs separately found a 73% increase in malicious open source packages over the same period, and Sonatype counted 454,600 new malicious packages across the major registries. The volume of validated findings a security team has to make a decision about had already outrun the review capacity built to absorb it, before either Gold Eagle or VulnHunter existed.

Building a better AI to find and validate more of those findings, faster, is not a mistake. It is necessary. But it is not the same project as deciding who is accountable for what happens to a finding once it is validated, on what timeline, with what documentation the organization can produce later if a regulator or a customer asks. That decision does not get made by a federal clearinghouse or an open sourced analyst tool, no matter how good either one is at its job. It gets made, or does not get made, in a boardroom, by people who have to decide in advance what "we knew and did not fix it in time" is going to cost them. A remediation program built around a documented timeline, 5 business days for a critical finding, 10 for high, 30 for everything else, against an industry average that still lags past 50+, is not a detection improvement. It is the accountability decision Gold Eagle and VulnHunter are both quietly assuming someone else already made.

That decision either exists in writing, with an owner and a clock attached to it, or it does not. Right now, for most organizations building or adopting the next generation of AI security tooling, it does not.

Frequently Asked Questions

What is VulnHunter?

VulnHunter is an open source, AI-powered security tool released by Capital One that starts from likely attack entry points, maps potential exploit chains, and attempts to disprove its own findings before surfacing them to human analysts, reducing false positives.

What is Gold Eagle?

Gold Eagle is a Treasury-led federal clearinghouse launched to coordinate AI-powered vulnerability discovery, validate findings, and connect agencies, companies, and open source maintainers with patching support.

Does Gold Eagle guarantee vulnerabilities get patched faster?

No. Gold Eagle can validate and route findings, but open questions remain about oversight, sensitive disclosure handling, and whether discoveries translate into completed patches on any particular timeline.

Why doesn't faster vulnerability detection solve the underlying problem?

Because detection and accountability are different layers. A better AI tool changes how quickly a validated finding arrives. It does not decide who is responsible for fixing it, on what schedule, or what documentation an organization can produce afterward.

What should a security leader do differently in response?

Document, in advance, who owns a validated finding and the remediation timeline they are held to, rather than waiting for a federal clearinghouse or an internal tool to make that decision by default.