COMMENTARY
For decades, finding a serious vulnerability in widely used open source software was specialized work. It took a skilled researcher weeks, sometimes months, to trace a flaw and responsibly bring it to a maintainer. That timeline has effectively collapsed. Advanced AI models can now produce vulnerability reports in hours, demolishing what a seasoned professional would have taken weeks to develop. That’s not hypothetical: it’s what the open source security community has watched happen since the fall of last year, as tools built on frontier models and strong open-weight models alike started turning out findings that are, frankly, good.
The trouble is that discovery was never the bottleneck. Fixing was. Patching, disclosure coordination, the upstream maintainer who has to understand a report, validate it, and ship a release — none of that has sped up at anywhere near the same rate. Some projects are using AI to combat the problem, like Valkey’s provenance guard or AIxCC winner Trail of Bits’ Buttercup finding vulns at DEF CON 2025.
Discovery in Hours, Remediation in Weeks
However, the structural mismatch remains: discovery measured in hours, remediation still measured in weeks and months. IBM’s Cost of a Data Breach Report 2026 puts a number on that gap. One in four malicious breaches last year were AI-enabled, up 56% over the prior year, and those breaches cost companies $6 million on average, roughly a million more than the overall breach average. The same research found only 18% of organizations are applying AI agents to vulnerability management, even as more than half already use agents for threat detection. Models are finding problems at maximum velocity while the humans supporting these projects are still sitting in the parking lot.
Adversaries already have comparably capable agents on their team, too, because the evidence says so. Open-weight models have closed much of the gap with the most expensive frontier systems, which is generally good for defenders: Openness lets you understand how a model was trained and steer it deliberately rather than trust a black box. However, that same openness lowers the floor for attackers, too. This is a current reality.
So, what needs to change?
Remediation and Prioritization
Remediation and prioritization have to become an engineering discipline. When the tens of thousands of lines of AI-discovered findings arrive in droves, treating each as an emergency is a recipe for burnout and bad triage. Projects need prearranged criteria for severity and exploitability, and reports need to reach maintainers validated and documented, not as a raw data dump that overwhelms a human reviewer.
Right now, a lot of that validation and routing work simply has no home: Multiple organizations independently scan the same obscure library, then each file separately without coordinating, multiplying the load on a maintainer who, on top of all of this, may be working on the project in their spare time. That’s the coordination gap efforts like Project Akrites are starting to fill, verifying findings, arming maintainers with context, and synchronizing disclosure, so a fix reaches everyone who depends on a package at the same moment it goes public. It’s one piece of a larger response the ecosystem needs and will only work alongside longer-running efforts to create best practices, financially support maintainers, and elevate secure-by-design work in the bouquet of security options for trusted AI.
Just as important is the human cost of the vulnerability gap. Maintainer burnout was a real problem before AI-generated reports started arriving in bulk, and a wave of well-meaning but uncoordinated disclosures makes it worse, not better. Part of a potential solution is identifying projects that are under-resourced or have lost their steward entirely, and connecting them with organizations able to provide sustained support. If more companies that benefit or build with open source software — which is most companies — look back upstream to support the origins of their products, we’d have fewer drowning maintainers.
All of this lands directly on the EU Cyber Resilience Act (CRA), a major requirement that has global repercussions for organizations selling to the EU market. Manufacturers have obligations to handle vulnerabilities responsibly and disclose them within tight timelines. AI-accelerated discovery and CRA deadlines are converging into a one-two punch for cybersecurity: more findings, arriving faster, against a regulatory clock that doesn’t move and comes with high fees. Getting vulnerability data out of upstream projects and into manufacturers’ hands quickly is one of the most concrete ways the ecosystem can help organizations meet those obligations, even though compliance remains each manufacturer’s own responsibility.
None of this is a call to slow down AI-assisted security research. It’s a call to build the coordination infrastructure, staffing, and standards that let the ecosystem actually absorb what that research produces. Finding problems faster only helps if the industry also gets faster, and more open, when it comes to fixing them. That work is underway, though countless individual maintainers are stuck holding the line. This effort needs more hands to close the gap.
Comments are closed