Skip to main content
onext technology
AI 16 November 2026 - 6 min read

Vulnerabilities AI finds that your team never gets round to fixing

If finding security flaws is now cheap, your team has to decide who filters them and how much time it sets aside to fix them.

Bernat López
Founder and CEO of onext
Three developers in a planning meeting: one points at a task board on a large screen; another, seen from behind, works on code on a laptop, and the third, in profile, looks at the screen thoughtfully

Monday planning. The review agent the team switched on a month ago has left another batch of security warnings on the board. Someone asks which ones are real. Nobody knows, because nobody has had time to look at them, and the delivery the client is waiting for already fills the whole iteration.

I propose to call this fixes with no slot. More warnings arrive every week, and the iteration plan is the same size it was a year ago.

Fixes with no slot: real security flaws, already found by a person or by a tool, that nobody has planned to fix because every iteration arrives full of what was promised.

The thesis

I am not saying you should switch off the scanner or wait until there is time for security; I am saying that, on the day your team switches on a tool that finds flaws, it has to decide who filters what arrives, by what criterion it is dismissed and what share of each iteration is set aside for fixing.

Filtering and fixing what arrives still takes people with judgement, and that work lands on the same team that has the next delivery promised. If nobody decides, the list grows and the fixes with no slot pile up.

What the data says

Anthropic publishes a dashboard of what its models have found in open source projects between November 2025 and November 2026. There are 37,966 candidate findings. External security firms have reviewed 6,376 and confirmed 6,009, or 94.2%. 5,087 have been reported to the people who maintain 680 projects, and the dashboard counts 629 already patched. Our own arithmetic: fewer than one in five candidates has gone through human review. The dashboard explains why: independent human triage and review is "the rate limiting step", and patches take a long time to create.

curl, an open source library for transferring data over the internet, went through this in two stages. In October 2025, its lead maintainer, Daniel Stenberg, wrote that two researchers using AI-based analysers had sent him more than 400 possible flaws. The first list led to about 50 fixes, with "remarkably few" complete false positives, and he still described it as "a fair pile of work".

In January 2026, Stenberg closed curl's bug bounty programme. Before 2025, more than 15% of reports were real vulnerabilities; from 2025, fewer than 5% ("not even one in twenty was real"). He wrote that the never-ending stream of empty submissions took "a serious mental toll" to manage. curl still accepts reports, with no money involved.

Veracode, which sells security tools, measures every year how long its customers take to fix flaws. In its 2025 report, the average time to fix a flaw was 252 days, up from 171 five years earlier, and half of organisations carried critical security debt: serious flaws left open for more than a year.

The sources have their limits. The dashboard is published by a model maker and measures open source, far from a company's own code; it also warns that a published patch does not guarantee it has been installed. curl is a single project, and its programme paid for each flaw, which attracted low-quality reports. Veracode measures its own customers. None of them measures a product team with a review agent, which is the case this article is about.

What it looks like in a team

  • The review agent comments on a possible flaw in every pull request (the request to merge a change) and the team learns to close it without reading it.
  • The scanner opens unestimated tickets, and in planning they move to the next iteration because they "can wait".
  • A real warning and a false one sit with the same priority, because nobody has decided which is which.
  • The person who can tell them apart is the same person who has to get Friday's delivery out.
  • A dependency with a published patch stays un-updated for months, because updating it breaks something and nobody has time to look.

Why it happens

What was promised to the client has a date and someone asking about it. An unexploited vulnerability is invisible: it triggers no calls and delays no demo. With that split, moving the fix to the next iteration is a reasonable decision every time it is made.

Part of the responsibility lies with management. The number of warnings the tool finds is celebrated in the meeting and nobody sets aside the time to fix them. The team gets two assignments at once, deliver and secure, and only one of them is measured.

What putting it off costs

Every unclassified warning lowers the value of the rest. When the list only grows, the team learns that its warnings change nothing and stops reading them, and the real flaw that arrives in the middle goes just as unnoticed.

Waiting also has a direct cost. With 252 days on average to fix a flaw at the companies Veracode measures, a list that grows faster than it empties turns every finding into debt. That debt shows up later, in an incident or in an upgrade that can no longer be done because you are three versions behind.

And the list fills from two sides. What scanners find is added to what the agent itself leaves behind when its code passes the tests and is still not secure.

What to decide before switching it on

If you do the work:

  • Classify every new warning within five working days at most: real, false positive or accepted risk with a written reason.
  • When you dismiss a false positive that keeps coming back, write the reason into the repository rules the agent reads, so the next review starts from there.

If you lead the team:

  • Before switching on the tool, write the severity criteria and what gets dismissed into the team's context, so warnings arrive already sorted.
  • Look every month at the time from when a flaw is confirmed to when it is fixed, next to the deliveries.

What Manel Marcos, our technical lead, would do:

  • An automated check that stops little, and stops well. Only a confirmed high or critical severity vulnerability in the code that this pull request changes blocks the merge (the integration of the change). Everything else goes into a queue that gets reviewed without stopping the team.
  • No warning without a proposed fix. If the agent finds a flaw, it also opens the change that fixes it, with its tests. Reviewing a fix costs far less than writing it from scratch.
  • A filter owner on rotation. Every iteration, one person decides which warnings are real and which are dismissed, with a written criterion. That way it does not always become the same bottleneck.
  • A fixed slot in every iteration. A stable part of the capacity goes to fixing, and the security queue is watched with the same metric as the rest of the work: how many come in and how many get closed.

I very often come across teams with a long list of technical debt, or with issues competing with new features without any prioritisation criterion. There, the scanner only makes the list longer.

Making room for what AI finds is a planning decision, and it is worth taking on the same day the scanner is switched on.

Frequently asked questions

What are fixes with no slot?

They are real security flaws that someone, a person or a tool, has already found and that nobody has planned to fix, because every iteration of the team arrives full of promised work. Giving them a place in the plan is a decision for whoever plans.

How many vulnerabilities does AI already find?

Anthropic's disclosure dashboard counts 37,966 candidate findings in open source projects between November 2025 and November 2026. External security firms reviewed 6,376 and confirmed 6,009, or 94.2%. The dashboard itself says human review is the step that sets the pace.

Why did curl close its bug bounty programme?

Because from 2025 fewer than one in twenty reports was a real flaw, compared with more than 15% before, and filtering that noise was wearing the team down. Daniel Stenberg announced it in January 2026. curl still accepts security reports, without payment.

How long does a company take to fix a security flaw?

In Veracode's 2025 report, the average time was 252 days, up from 171 five years earlier, and half of organisations had serious flaws open for more than a year. Veracode sells security tools and measures its own customers.

What is vulnerability triage?

It means classifying every warning that arrives: whether it is a real flaw, a false positive or a risk accepted with a written reason. It is worth doing within a few days and with criteria written into the team's context, so the list does not grow without anyone understanding it.

How much capacity should be set aside for fixing?

There is no figure that works for every team. The practical approach is to set a percentage of each iteration, keep it for three iterations and adjust it with what has come in and what has been fixed in that time.

Sources

Bernat López
Written by
Bernat López
Founder and CEO of onext

Bernat López is founder and CEO of onext, an AI boutique. He helps development and product teams work with AI with method —specification before building, automated checks and Spec-Driven Development— and applies to his own company what he proposes: onext runs on its own agentic system.

LinkedIn →

How many security warnings does your team have unclassified?

We look at it with you in the diagnostic, which is free.

See how we work

What we build stays with your team.