Every security program I have ever seen is organized around scarcity of findings. Scanners are scheduled like audits. Penetration tests are annual events with waiting lists. Findings are rationed into a queue because everyone downstream, the engineers who have to fix them, is the real constraint, and the whole system quietly agrees to never produce findings faster than the fix queue can absorb them.
Frontier models just ended the scarcity. The program has not caught up.
The cleanest public receipts are Anthropic's. Project Glasswing put the Claude Mythos model to work across roughly 50 partner organizations, scanning real codebases, and surfaced more than 10,000 high and critical vulnerabilities.[1] That number got the headlines. The number underneath it is the one that should reorganize your program: per Anthropic, more than 99% of the candidate findings were still unpatched.[1] The best-resourced security organizations on earth, handed verified findings by the lab that built the model, could not fix them at anything like the rate they arrived.
Microsoft's numbers say the capability is not one lab's party trick. Its MAI Cyber 2 Flash model scores 92.3% on CyberGym-E2E, a public benchmark of real vulnerability discovery and exploitation tasks, at a small fraction of the cost per task of the frontier systems it edges out.[2] Specialized cyber models, frontier generalists, open weights improving behind them: the direction is uniform. Finding vulnerabilities in code is becoming a commodity input, priced per scan and falling.
So the bottleneck moved, permanently, to the two places it was always going to end up: deciding what matters, and fixing it.
Deciding what matters is a verification problem before it is a prioritization problem. A model scan at estate scale produces candidates, not conclusions. Some fraction are real and exploitable, some are real and unreachable, some are wrong. The programs that work will treat the scan as the first pass of a pipeline whose second pass is adversarial: attempt the exploit path, in a replica or a range, and let the result sort the queue. Severity scores do not sort a queue of ten thousand. Exploitability does.
Fixing is a throughput problem, and it has a new shape. The old unit of remediation was a ticket a human engineer picked up cold: reproduce the finding, understand the code, write the fix, test it, ship it. That unit does not scale to model-speed finding, and no hiring plan closes the gap. The new unit is an assisted fix: an AI system drafts the patch with the finding, the exploit test, and the surrounding code in context; a human reviews a diff instead of starting from a blank editor; the exploit test runs again to confirm the path is closed. Tools like Copilot autofix are the early commercial shape of this. The pattern matters more than the product: generation is cheap, so the human's job concentrates where it was always most valuable, judgment on a concrete proposed change.
This changes what a scanning program should look like, concretely:
There is a failure mode waiting here, and it is the most likely outcome for most organizations: scan-and-shelve. Leadership buys the frontier scan, the number is enormous, the number is briefed, and the queue rots at 99% unpatched while everyone involved calls it a visibility win. Glasswing's partners are the best case and they are the warning. Visibility was never the mission. A vulnerability that is found, verified, and left open is worse than an unknown one in exactly one way: now the negligence is documented.
The next part takes up the loop that closes this one: the second scan, and why a fix should count only when the exploit fails.
[1] Anthropic, Project Glasswing (April 2026): Claude Mythos Preview evaluated across roughly 50 partner organizations; 10,000+ high and critical vulnerabilities surfaced; Anthropic reports more than 99% of candidate findings unpatched at publication. Partner-reported program; figures are Anthropic's.
[2] Microsoft, MAI Cyber 2 Flash (November 2026): 92.3% on CyberGym-E2E (Berkeley RDI; results reported on the 813-task set after exclusion of 92 tasks overlapping training projects, per Microsoft's integrity note), ahead of GPT-5.6 Cyber at 91.5% on the same reporting, at under 2% of the cost per task. Model is closed and ships inside Microsoft security products only.