AI found ten thousand bugs. The breach used one from last year.
Model-assisted bug hunting is flooding disclosure queues with real, mostly minor findings. Exploitation still runs on n-days, and the flood is making them harder to see.
By June Abernathy, Wire Editor, Dispatches4 min read
For most of the history of vulnerability disclosure, the constraint was finding the bugs. This year, in conversations with maintainers, bug bounty triagers, and the people who run vendor security response teams, the constraint has moved. The bugs arrive on their own now, by the thousand, mostly real, mostly minor, and almost all of them written up by a model.
The flood
AI-assisted bug finding works. Give a model a codebase and a fuzzer, or point an agent at a web application with a scope document, and it produces findings: real memory-safety issues in parsers, real authorization gaps in rarely used endpoints, real injection bugs in error handlers. One open-source project told us its security inbox went from 30 reports a year to 1,100 in nine months. A triager at a bounty platform described the queue as "ninety percent genuine, five percent duplicates of each other, and five percent genuinely important, in no particular order."
The problem is not that the reports are false. Most are not. The problem is that discovery has become cheap while the expensive parts, triage, fix, test, and release, remain human, and a report that takes a model four minutes to produce takes a maintainer four hours to confirm. Projects are rate-limiting their disclosure forms. Some have closed them. The important bug sits in the same queue as the 1,099 others, and it looks the same.
The breach
Meanwhile, the incidents we reviewed this quarter look like the incidents from every previous quarter. An edge appliance, a vulnerability patched eleven months ago, a working exploit published within a week of the patch, mass scanning for exposed instances, and an organization that had not got to it yet. Nobody involved used a zero-day. Nobody needed to.
An n-day is a vulnerability for which a patch has existed for n days. It is where attackers live, and it is where they still live, because the patch gap for VPN concentrators, firewalls, file-transfer appliances, and mail gateways is still measured in weeks to months, and because once a patch is published, comparing old and new code to build an exploit is itself a task models are good at. The same tools flooding defenders' queues are shortening the attacker's time from patch to exploit.
"Our vulnerability count went up tenfold this year. Our list of things actually being exploited against us is the same length it was, and it is the same list." — Head of security at a logistics company
Why the asymmetry persists
- Discovery scales; remediation does not. A model can find bugs in parallel. A release still ships one version at a time, and an appliance still needs a maintenance window.
- Volume hides severity. Triage teams built for dozens of reports a year do not find the one that matters in a queue of a thousand. The signal is real; it is simply buried.
- Exploitation follows exposure, not novelty. Attackers scan for what is exposed and old. New findings rarely matter to them until a patch publishes and the gap opens.
- Inventory is still incomplete. You cannot patch what you do not know you have, and no amount of discovery in your own code helps with the forgotten appliance in a subsidiary's rack.
How to prioritize
- Use the catalogs of known exploited vulnerabilities maintained by national cyber agencies as the primary patch queue. These are in active use. The report in your inbox, however real, usually is not.
- Patch internet-facing systems first, with a target measured in days. Track mean time to patch for exposed systems and report it as a risk metric; it predicts breach better than any count of open findings.
- Maintain an external attack-surface inventory and scan it yourself, because the attacker's model already has.
- Use AI for triage, not just discovery. The same tools that generate reports can deduplicate them, cluster them, reproduce them, and rank them by reachability and exposure. A queue of a thousand findings is a tooling problem; treat it as one.
- Ask reporters, and your own tools, for a working reproduction and a reachability argument before a finding enters the fix queue. Volume without proof is noise with a severity score.
The honest framing
Zero-days matter, and some organizations are genuinely targeted by them. But the question after each headline about a model finding a thousand bugs is unchanged: is the fix out, and have we applied it? If the answer to the first is yes and the second is no, you do not have a discovery problem. You have an n-day problem, and it is entirely yours to solve.
About the writer
Wire Editor, Dispatches
June edits the wire desk, where the stories are short and the patch gaps are long. She specializes in supply chains, vulnerabilities, and the maintainers nobody pays.