The story everyone is telling about AI and vulnerability disclosure is that generative models made it free to file a report, so maintainers are drowning in confident, plausible nonsense. Curl's public frustration with AI-written reports made the term "slop" stick.
The numbers do not support the strong version of that story.
What HackerOne published
Alongside its h1 Validation launch in April 2026, HackerOne released figures from its own platform:
| Metric | Figure |
|---|---|
| Submission growth | +76% year over year |
| Peak | Record high, March 2026 |
| Confirmed exploitable | ~25% — unchanged through the surge |
| Critical + high severity | 32%, up from a 26–28% baseline |
| Remediation improvement | +19% year over year |
The third row is the one that breaks the narrative. If AI were mostly producing garbage, the valid rate would fall as volume rose — that is what dilution looks like. It held.
And the fourth row goes the other way entirely: the proportion of critical and high findings went up, from a long-run baseline of 26–28% to 32%.
So the surge is not disproportionately junk. A quarter of a much larger number is still a quarter, and a larger share of it is serious.
The actual problem is arithmetic
Put rows one and five together.
Submissions up 76%. Remediation up 19%.
That gap is the whole story. AI reduced the cost of finding and did nothing to the cost of fixing. A vulnerability programme is a pipeline, and one stage just got four times faster relative to the one after it.
The consequence is not a quality crisis. It is a backlog crisis — more real, more severe findings arriving than the engineering organisation on the other end can absorb.
That is a materially different problem, and it has different fixes. Slop is solved by filtering. Backlog is solved by prioritisation, by fixing classes rather than instances, and by deciding — explicitly — which real vulnerabilities you are choosing not to fix this quarter.
Where the slop story is still right
Two things save the popular narrative from being simply wrong.
Triage cost is real and it is not evenly distributed. HackerOne is a platform with paid triage. An unpaid open-source maintainer receiving AI-generated reports has no such buffer, and their valid rate is not HackerOne's valid rate. Curl's experience is not contradicted by platform-wide statistics; it is a different population.
A 25% valid rate means 75% is not valid, and reviewing those still costs analyst hours. The rate holding steady is good news about quality and neutral news about workload.
The conflict of interest, stated plainly
These figures come from a vendor press release announcing a product that solves the problem the figures describe. HackerOne's Chief Product Officer Nidhi Aggarwal is quoted saying AI is accelerating both the volume and sophistication of vulnerabilities, and that h1 Validation helps organisations reduce the time from find to fix.
That does not make the numbers wrong. HackerOne has better platform data than anyone, and a vendor with a book of business does not benefit from inflating a valid rate — if anything the marketing incentive runs the other way.
But it means: unaudited, self-reported, and released in service of a launch. Treat accordingly.
One further caution. A specific submission count for March 2026 — 46,947 — is in circulation. It does not appear in the press release we read, and we could not trace it to a primary HackerOne document. Use the 76% figure, which is sourced, rather than the absolute number, which is not.
What this changes for a security team
- Stop budgeting triage as a fixed cost. If submissions are up 76% and your intake process is unchanged, the queue is where your risk now lives.
- Measure your own valid rate. If it is far below 25%, you have a scope or a signal-to-noise problem specific to you, not an industry one.
- Watch the severity mix, not just the count. 32% critical-and-high on a much larger base is a real change in exposure.
- Fix classes, not instances. The only response that scales against a 76% increase is remediating the pattern that produced the finding.
- Be honest about the remediation number. 19% is what your engineering capacity actually grew by. Everything above that accumulates.
The uncomfortable conclusion is that the "AI slop" framing has been comforting. It says the extra work is fake and can be filtered away. The data says most of the extra work is real, more of it is serious than before, and it is arriving faster than anyone can fix it.
