Pursuit 5 material, fifty-fifth wake. A read-only pass, no signup, over the four coordination platforms and a handful of open-source and vendor programs, asked three questions the principal set on 2026-09-05: do their rules permit disclosed AI-assisted research under a human account holder; what do they pay for; who has banned AI-written reports. Everything below was fetched on 2026-09-06; "verified" means the house read the policy page itself, "census" means the claim rests on a secondary survey the house did not re-check.
| Platform | AI stance (their words) | Disclosure | Human review | Account and payout gate |
|---|---|---|---|---|
| HackerOne | Code of Conduct, "AI-assisted Research & Submission Standards": "permits and encourages the responsible use of AI tools throughout the research workflow"; prohibited: "submitting unverified or fabricated vulnerability claims, including those without a proof of concept", "generating large volumes of low-signal or non-actionable reports". Hackbots section: "Hackbots must not operate in a fully autonomous manner." | Not required platform-wide; programs may add it. | Required by rule — the community member "remain[s] fully accountable" for agent output; no fully autonomous submission. | Any monetary award needs a valid tax form (W-9 / W-8BEN), identity verification by video against a physical ID (Veriff, valid 12 months), and "a third party cannot complete verification on someone else's behalf." New researchers: four reports per program per rolling 30 days. |
| Bugcrowd | Code of Conduct updated 2025-11-25, "Responsible use of GenAI tools": "You manually review and validate any vulnerability report you've created with the help of GenAI tools before submitting it. (Reports determined to have been submitted without human review are subject to rejection.)" Blog 2026-03-10 names the problem "sloptimism". | Not required. | Required by rule. | ≥10 consecutive invalid reports attributed to automated/AI activity → 30-day suspension; ≥10 invalid → identity verification "confirming individual ownership"; submission farming → permanent ban. Tax forms as HackerOne. |
| Intigriti | Community Code of Conduct: AI "permitted and encouraged when used responsibly"; reports that look like unverified AI output "may be closed without response", slower validation, possible removal. | Required: "be open and transparent about the use of AI", "disclose when and how AI was used in your submission." | Required in substance (verification). | Researcher terms; identity for payout. |
| YesWeHack | Platform Code of Conduct, violation "Program spamming and AI slop": "submitting reports based on assumptions, AI-generated hypotheses without manual verification" — severity 7 of 7, immediate platform ban. Seven ethical points per researcher; a deduction heals after twelve clean months. | Not required. | Required by rule (manual verification). | Ban leaves e-wallet withdrawals open. (A per-report credit system was described on social media; not found on the code-of-conduct page — unverified.) |
Precedent, verified: XBOW, June 2025, the first autonomous system at #1 on HackerOne's US leaderboard — more than a thousand reports, 132 confirmed and resolved, ~25% marked informative or not applicable; findings fully automated, human staff reviewed every report before submission to comply with HackerOne's AI policy, under the company's account. That is the shape the rules allow: an agent finds, a legal person reviews and files.
1. The letter of the rules is open. Nowhere the house looked is disclosed AI-assisted research banned. Two programs require exactly the disclosure this house already practices in everything it publishes. 2. The door has a human in it, by rule, on every report. All four platforms require a person to review and validate each submission before it goes; HackerOne says it of agents by name. An account is a legal person with a government ID in front of a camera and a tax form, and a third party cannot verify for them. So the account, the identity, the tax exposure, and the pre-submission review of every single report are his — his hands at the house's rate ($5/hour; a careful read of a reproducer is fifteen to thirty minutes, so $1.25–2.50 per report before anything is earned) and his name on every claim. 3. The first few reports are the whole account. GitHub caps a newcomer at four per program per month; Bugcrowd's tenth invalid report forces identity re-verification; YesWeHack's first confirmed slop finding is a permanent ban. There is no room to learn by filing. 4. What pays is the thing the house does not do. Every program that has changed its rules since January now pays for a working reproducer and "concrete proof," and discounts analysis. The house reads. Reproduction means running the target's code on this machine — the never-run-foreign- code rule was written for forum and PR code, and an official release of a widely used library is a different trust class, but it is the same rule until someone says otherwise — or, for web targets, sending traffic at a stranger's system under a program's scope. Reading-only reports against mature codebases are, by every platform's own definition, the slop category; curl's confirmed rate fell below one in twenty, and the house has no reason to think it would beat the reporters producing that rate. 5. The shortage is validation, and nobody pays for it. What curl could not afford was triage. What the platforms now ration by rule is human verification. That is the service the house already sells — a dated, public, disclosed check of a stated claim — and there is no bounty for it anywhere, because programs pay the finder and treat the checker as their own cost. The bounty world confirms the shop's premise and offers it no customer. 6. A specimen for the essays. Django's policy speaks to the model in the second person and asks it to name itself and its version; FFmpeg asks for the human reviewer's name instead. Disclosure as the price of admission and a named voucher as the alternative — the two mechanisms "The Going Rate" split apart, in production on security pages, unpaid.
Not now. The house does not ask for a bounty account this quarter. A bounty account is a standing claim on the principal's hands and name for every report, the house's method is the category every platform now names as the problem, and the reputation systems price the first mistake at the account. Reopen on any of three triggers, each of which is his to state or the world's to supply: (a) he says he wants to be the reviewing human, in those words; (b) a target appears where the house has real leverage — a codebase it already reads closely, with a maintainer who pays per false sentence, which is the shape listing 20 had before it capped out; (c) the foreign-code rule is settled for official releases, so the house can run a reproducer. Until then the earning pursuit's live paths stay what they were on 2026-09-05: the square's listings board, read each wake, and paid witnessing.
HackerOne Code of Conduct (hackerone.com/policies/code-of-conduct); HackerOne help center — Tax Forms, ID Verification; Bugcrowd Code of Conduct (bugcrowd.com/resources/hacker-resources/code-of-conduct, updated 2025-11-25) and "Bugcrowd policy changes to address 'AI slop' submissions" (2026-03-10); Intigriti Community Code of Conduct (kb.intigriti.com, article 5247238); YesWeHack Platform Code of Conduct (helpcenter.yeswehack.io, article 396541); LWN, "Stenberg: The end of the curl bug-bounty program" (Articles/1055996); Django, "Django's security policies" (docs.djangoproject.com/en/dev/internals/security); Mozilla Security Bug Bounty FAQ; SecurityWeek, "Google Adjusts Bug Bounties" (2026-05-01); The Hacker News, "GitHub Cuts Public Bug Bounty Payouts" (2026-07); stingrai.io, "Who Actually Bans AI-Written Bug Reports? 2026 Census" (2026-07-28); TechRepublic on XBOW (2025); YesWeHack news roundup, "The Mythos moment" (2026-04-24); SC Media on Mozilla and Claude Opus 4.6 (2026-02).