Running a Bug Bounty When AI Writes the Reports: PHP’s 106 Closed, curl’s Exit

AI has made security reports cheap to produce and just as slow to check. PHP closed 106 reports since July 1, curl ended its paid program, and an AI tester topped a HackerOne leaderboard. Here is what that means for a startup deciding how to take reports.

On Oct. 14, the PHP security team published a plain account of what its inbox looks like now: since July 1 it had published 15 CVEs and closed 106 reports that were not real vulnerabilities, were duplicates or did not reproduce. A week later PHP shipped fixes for the critical PHP2Shell flaws; The Register reported that the three firms behind those finds “have said they used AI models to find the vulnerabilities.” For a founder weighing whether to open a program, the bug bounty AI problem cuts both ways: more real findings, and far more noise to sort through.

Where it stands: The PHP team says “AI tools have made security reports much cheaper to write over the last year.” The curl project stopped paying for reports on Jan. 31, 2026. HackerOne paused submissions to its Internet Bug Bounty in April 2026, according to Privacy Guides.

What does the bug bounty AI flood look like in numbers?

The clearest data comes from maintainers who published their own counts. None of these figures measure the same thing, so they should not be added together, but each shows the same pressure from a different seat.

Who reported it What they counted
PHP security team (since July 1) 15 CVEs published; 106 reports closed
curl (2025) Confirmed rate fell to below 5%, from north of 15%
XBOW on HackerOne (June 2025) Nearly 1,060 submitted; 130 resolved; 208 duplicates
HackerOne’s annual report Autonomous agents filed 560+ valid reports

A CVE is a public ID assigned to a confirmed security flaw. A bug bounty pays outside researchers for valid findings; a vulnerability disclosure program, or VDP, gives them a safe way to report without promising money. Both depend on someone reading what comes in.

Why are AI-written reports so expensive to triage?

The cost sits with the receiver, not the sender. The PHP team put it directly: “Raw model output with a crash that doesn’t reproduce takes us just as long to rule out as a real bug takes to confirm, and we’ll have to close it.”

HackerOne made the same point when it joined OpenAI’s Patch the Planet program, writing that the problem “has only grown as automated tooling makes it cheaper to file a report than to confirm one.” For a startup with one or two engineers who know the codebase deeply, every bogus report pulls exactly those people off product work.

That asymmetry is why a bounty is a staffing decision as much as a security one. Before announcing rewards, a founder should know who reads reports, how fast, and who has authority to close one.

What happened when curl ended its bug bounty?

curl’s lead developer, Daniel Stenberg, closed the program on January 31, 2026, after years of results. He wrote that it produced “87 confirmed vulnerabilities and over 100,000 USD paid as rewards to researchers.” Then the mix changed. “Previous years we have had a rate of somewhere north of 15% of the submissions ending up confirmed vulnerabilities. Starting 2025, the confirmed-rate plummeted to below 5%.”

He blamed “an explosion in AI slop reports” plus lower quality in the rest. curl still takes security reports; it just stopped paying. “We no longer offer any monetary rewards for security reports – no matter which severity.”

Stenberg is not anti-AI, and that matters for founders reading his post as a verdict. In May 2026 he wrote: “We also see a high volume of high quality security reports flooding in : security researchers now use AI extensively and effectively.” Ending the payments did not end the reports, and he still credits researchers who use AI well.

What does XBOW topping HackerOne tell a startup?

It shows automated testers can now compete at the top of a public program. XBOW wrote in June 2025: “For the first time in bug bounty history, an autonomous penetration tester has reached the top spot on the US leaderboard.”

The fine print is useful for planning. XBOW says it “submitted nearly 1,060 vulnerabilities,” that its own team “reviewed them pre-submission to comply with HackerOne’s policy on automated tools,” and that “208 were marked as duplicates, 209 as informative and 36 as not applicable.” Even an AI operator with human review in the loop produced hundreds of reports that programs marked duplicate or informative rather than resolved.

HackerOne’s own report, reproduced by Disaster Recovery Journal, says “70% of surveyed researchers now use AI tools in their workflow.” A startup opening a program should assume most incoming reports have touched an AI tool somewhere and write rules for that, rather than trying to ban it.

How can founders set triage rules for AI-assisted reports?

Borrow the rules maintainers already wrote, and publish them before the first report arrives. The PHP team’s policy is a good template because it does not reject AI; it rejects unverified output: reports are “welcome, as long as you did what we’d ask of anyone: reproduce the issue on a PHP build , check it against the policy, and understand what you’re sending.”

Turned into a startup policy, that suggests a few concrete choices:

  • Require reproduction. Ask for steps that work against your current production version and close anything that does not reproduce.
  • Define scope tightly. List which domains, apps and APIs count, and which classes of report you do not pay for.
  • Group by root cause. The PHP Foundation says AI helps it “group reports sharing the same root cause.” Pay once per root cause, not once per report.
  • Keep a human on the decision. PHP’s line is “The models don’t make the calls, though.” Use AI to sort the queue if you like; a named person still decides.
  • Start without money. A disclosure policy with no cash, as curl now runs, still gives honest researchers a channel.

Rewards can come later, once someone owns the queue and you know your monthly volume. Our look at AI security scanning for startup codebases covers the other route: finding the bugs yourself before outsiders do. And when a report does turn into a fix, our patch day runbook from Oct. 21 covers shipping it.

Will the volume keep rising?

The people closest to it expect so. The PHP Foundation wrote on Oct. 21: “Tools now find bugs faster than ever, for defenders and attackers alike, and the number of reports will keep growing.” A bipartisan House letter in May 2026 made the same argument at national scale: “AI can empower defenders to discover a multitude of serious vulnerabilities, but the corresponding disclosure, validation, patching, and deployment efforts may struggle to keep pace.”

Large platforms are adjusting too. Privacy Guides reported that HackerOne said it was “pausing submissions” to the Internet Bug Bounty because AI was overwhelming recipients, and HackerOne wrote of its Patch the Planet work: “We don’t count success in report volume.” The PHP team, a week before the PHP2Shell fixes shipped, counted 15 CVEs against 106 closed reports.

Frequently Asked Questions

Should a startup run a bug bounty when AI makes reports cheap to write?

It depends on whether the team can triage what arrives. The PHP security team says AI made reports much cheaper to write, and curl ended its paid program after its confirmed rate fell below 5% in 2025. A small company can start with a no-payment disclosure policy and add rewards once it has someone to review reports.

What is the difference between a bug bounty and a vulnerability disclosure program?

A vulnerability disclosure program, or VDP, is a published way for outsiders to report security flaws, usually without payment. A bug bounty adds cash rewards for valid findings. After January 31, 2026, curl kept accepting security reports but stopped paying for them.

How do you filter AI-generated bug bounty reports?

Ask for what the PHP security team asks for: reproduce the issue on a current build, check it against the published policy, and show you understand what you are sending. Reports that cannot be reproduced get closed. The PHP team also uses AI to group reports that share a root cause, with a person making every call.

Did an AI really top the HackerOne leaderboard?

XBOW says its autonomous penetration tester reached the top of HackerOne’s US leaderboard in June 2025 after submitting nearly 1,060 vulnerabilities. XBOW says its own security team reviewed findings before submission to comply with HackerOne’s policy on automated tools, and its figures include 208 duplicates and 209 marked informative.

Is AI making bug bounty reports worse or better?

Both views appear in the record. curl’s Daniel Stenberg described an explosion in AI slop in his program’s final period, yet in May 2026 wrote that he also sees a high volume of high-quality reports from researchers who use AI effectively.

Sources

This article is general information for founders, not legal or security-consulting advice. Incident details are as reported by the sources cited on the dates shown and can change; for breach-notification duties, talk to counsel.