Decentralised News · 2027 planning series · Reviewed 21 November 2026
What Matters
An AI-assisted security report can be useful without establishing that an autonomous agent earned a bounty. Permission to research, evidence of a vulnerability, programme eligibility, an award decision and receipt of money are separate facts. A credible payout claim needs the records connecting them.
This guide gives six checks for that chain. It does not rank alleged autonomous payouts, test an exploit or estimate how much a model can earn. Its original contribution is a small evidence method that preserves the difference between an attractive finding, an accepted report and a documented payment. All numerical examples are synthetic.
The Signal: assistance is documented; entitlement still depends on the programme
HackerOne’s Hai for Hackers documentation describes help with report writing, interpretation of supplied material and communication. That is evidence of an assisted reporting workflow. It is not a record of autonomous discovery or a promise that the resulting report will receive a reward.
Immunefi’s current platform rules address scope, required impact evidence, testing restrictions and incomplete AI-generated or scanner reports. Individual programmes add their own terms. The editorial question is therefore precise: what does this programme permit, what evidence was submitted and what outcome actually followed?
DN Alpha Thesis
The strongest bounty story follows one report through permission, evidence, evaluation and payment. Attribution to AI should be a separate line inside that record. Keeping the lines separate makes it possible to recognize useful automation without turning an assisted report into an unsupported claim of autonomous income.
1. Establish the applicable permission before evaluating the finding
Start with the programme that was active when the research occurred. Record the owner, programme reference, relevant version or capture date and the permitted research environment. A public repository, a visible contract address or the ability to call an endpoint does not by itself establish the programme’s authorization boundary.
The relevant rules may exist at more than one level. Platform rules can govern conduct across programmes, while a particular programme specifies assets, impacts and special conditions. Read both and preserve any conflict as unresolved. A generic article written years earlier should not silently replace a current programme rule, and a favourable sentence should not be extracted from conditions that materially limit it.
For a concrete example, the Money on Chain / RIF on Chain programme explicitly welcomes evidence-based AI assistance while expecting meaningful human follow-up and a proof of concept. Those are that programme’s published expectations. They do not establish a universal policy for every platform or authorize live testing outside the listed scope.
DN recommends a permission record with three separate statements: the research action contemplated, the source believed to permit it and the limits that remain. If the relationship is unclear, resolve it through the programme’s approved process before proceeding. The worksheet only stores reference aliases and declared states; it does not interpret a contract, authorize testing or contact a target.
This distinction matters when a model suggests an additional test. A technically interesting suggestion can exceed the agreed environment or touch another party’s systems. The responsible researcher must preserve the scope boundary. Automation should not treat a failure to obtain a result as permission to widen the target.
2. Tie the report to an in-scope asset and a specific impact
A programme may accept some assets and some consequences while excluding others. A report therefore needs more than a project name and a severity label. Identify the actual artifact or service, the relevant version and the claimed property that fails. Connect the consequence to the programme’s accepted impact category.
Keep identity and impact evidence separate. The record naming a deployed version answers which object is being discussed. The evidence showing an unauthorized state change, incorrect calculation or other failure answers a different question. A label such as “critical” cannot substitute for either record, and an impressive amount of value associated with a project does not establish the amount affected by the particular finding.
Immunefi’s rules prohibit misrepresenting assets, severity and impacts. They also address required reproduction material. The practical consequence is to retain the original evidence and the limits of the claim, including prerequisites that have not been shown. A report can be technically plausible while its claimed real-world impact remains unestablished.
DN’s existing smart-contract audit evidence guide follows the artifact, property, reproduction, repair and retest chain in detail. The additional bounty question is whether that finding falls within this programme’s terms. An audit conclusion and a reward decision can differ without either necessarily being inconsistent: they may be answering different questions.
Preserve negative and uncertain findings. If a condition could not be reproduced on the named version, record that outcome and the attempted environment. If another asset is affected, treat it as a separate scope question. Do not broaden the report’s wording to conceal a mismatch between the evidence and the permitted target.
3. State what the AI did and what another person established
The phrase “AI found the bug” can cover very different workflows. A model might generate an initial hypothesis, suggest a test, summarize an existing warning, improve a report’s language or perform an authorized search under an operator’s direction. Those contributions should not be flattened into one autonomy label.
Use a task-level provenance record. Note which input the model received, the model or tool version where relevant, the output that contributed to the finding and the subsequent human or independent validation. Preserve the distinction between original output and later editing. A fluent final report does not reveal which part of the reasoning was correct on the first attempt.
A practical attribution statement might say that AI assisted with organizing notes while a researcher identified the issue and reproduced it in an authorized environment. Another record might show a model-generated candidate that was later validated by a human. Both can describe useful AI work. Neither should be promoted to autonomous end-to-end discovery unless the evidence actually supports that stronger statement.
HackerOne documents AI support inside its reporting environment, including constraints on using reports when a customer has disabled the feature. This supports a further operating question: was the report material permitted in the chosen tool? A researcher should not move confidential target information into an unrelated model service merely because it offers better prose.
For a public case study, use only material that may be disclosed under the applicable programme. Keep protected logs and sensitive technical details in the authorized evidence channel. An attribution summary can explain the division of work without publishing credentials, customer data or an operational exploit. The local worksheet is deliberately limited to aliases and classifications.
4. Preserve reproducibility, triage and the unresolved questions
A submitted report begins a review process; it does not conclude one. The recipient may need clarification, identify a duplicate, disagree on impact or request additional permitted evidence. Record those events without treating the first acknowledgment as a finding of validity.
HackerOne’s triage documentation describes automated assessment and human analyst involvement. It distinguishes initial classification and recommendations from analyst decisions about validity, severity and bounty. For the reader evaluating an “agent earned money” headline, a message generated during intake establishes only the action that message actually reports.
The reproducibility record should identify the environment, artifact, expected property, observed result and evidence location. The public article need not contain the sensitive reproduction procedure. A reviewer should nevertheless be able to locate an authorized record that supports the stated outcome. If the evidence cannot be shared publicly, say which party evaluated it and what remains unavailable for independent assessment.
Duplicate status deserves its own field. A valid technical issue may already be known under the programme’s rules. Conversely, a claim that something is a duplicate requires a programme-specific evaluation. The worksheet does not decide novelty from a report title or timestamp. It preserves the reported decision and the source behind it.
Keep the chronology intact when the conclusion changes. A model hypothesis can become a rejected report; a disputed impact can later be accepted; a repair can change the relevant version. A final success story that omits the earlier uncertainties makes the process look more autonomous and repeatable than the record establishes.
5. Distinguish an award from a payment received
A reward table states possible programme terms. An award decision concerns a particular report. A payment instruction starts a transfer process. A receipt records money reaching the intended recipient. These are separate evidence levels, and they should remain separate in a financial or editorial claim.
HackerOne’s payment guidance describes payment arrangements and preferences. Its mediation guidance distinguishes a bounty disagreement from a payment that did not go through. Those documented categories illustrate why an award and a received amount cannot automatically be treated as the same observation.
Use a dated award reference and a separate payment reference. Identify the currency or token, recipient relationship and amount relevant to each. If the reward was split among collaborators, report the allocation basis rather than attributing the whole gross award to one model or operator. If a public payment record is insufficient to connect the transfer to a particular report, retain the missing link.
A token transfer also needs a valuation boundary before it becomes a fiat-income statement. The amount received, a later sale and an assumed conversion price are different records. This article’s local cash calculation uses one reader-declared unit and does not convert tokens, calculate tax or estimate realizable market value.
HackerOne’s statistics documentation defines reward metrics using payments distributed during the selected period. DN’s editorial application is to choose a matching denominator and clock. “Reports submitted this month” and “rewards received this month” may concern different sets of work. Combining them can create a misleading apparent success rate.
6. Put costs and disclosure conditions beside the payout
A paid bounty can exceed the cost of one successful model call while failing to cover the wider research programme. Count the boundary that matches the claim: unsuccessful candidate reports, compute, authorized environments, review effort, tools, payment costs and any other included expenses. Explain which are observed cash costs and which are separately estimated labour or overhead.
Do not convert an incomplete cost record into a precise profit margin. If only the successful attempt’s model bill is known, the supported statement is narrow. It does not describe the cost of discovering the issue, the total expense of the team or the expected return from repeating the process. A ranking of “most profitable autonomous agents” would need comparable boundaries and substantially more evidence.
The local worksheet calculates received amount minus entered included costs only when the user declares that a payment is documented, supplies its reference and identifies a complete cost boundary for that calculation. The result is a scoped cash difference. It is not a tax result, a model earnings forecast or a measure of the value of the vulnerability to the project.
Publication conditions are part of the record. Programme rules can restrict disclosure of a report’s existence, details and supporting material. Record the disclosure status and source before preparing a case study. Receiving money does not, by itself, establish that the researcher may publish every associated log, screenshot or technical detail.
The useful public claim may remain modest: an authorized AI-assisted report was accepted, a specified award was announced and a specified amount was received according to named evidence. If one of those stages is missing, state the missing stage. That approach makes real progress visible without inventing a list of autonomous payouts.
A worked example: an award headline outruns the available cash record
Consider an invented programme, REPORT-A and a fictional researcher. The programme and AI-assistance rules have reference aliases. A controlled reproduction record exists. The model helped organize the initial candidate, and a human performed the permitted reproduction and follow-up. An award notice states 5,000 currency units, but no payment receipt has yet been recorded.
The researcher has recorded 600 units of included cash costs. The worksheet displays the declared award separately and withholds the received-minus-cost calculation. It does not substitute the award amount for received cash. That is the decision-relevant result: the announcement supports a claim about an award, while the actual receipt remains unresolved.
In a second synthetic state, the researcher enters a payment reference and 4,800 units received in the same unit, with the 600-unit cost boundary declared complete. The conditional cash difference becomes 4,200. The 200-unit difference from the award must be explained by actual records; the worksheet does not guess that it represents a fee, a split or a missing transfer.
| Claim | Record needed | What it cannot establish alone |
|---|---|---|
| AI contributed | Task-level provenance | Autonomous discovery or programme permission |
| Finding accepted | Programme evaluation reference | An award or receipt |
| Award made | Report-linked award decision | Received money |
| Money received | Payment evidence and report linkage | Complete programme profit or repeatable earnings |
DN Bounty Claim Evidence Record
Use one record per claim and programme version. Enter aliases rather than report contents, personal information or exploit material. The default scenario is the unpaid fictional award above. Evidence states remain user declarations. The tool does not decide scope, validate a vulnerability, authenticate payment or authorize disclosure.
DN Bounty Claim Evidence Record
Change the illustrative assumptions below. This worksheet runs locally in your browser. It records your declarations; it does not verify evidence, connect to accounts, submit instructions or grant approval.
The initial numbers are synthetic. Unknown evidence remains unresolved. Downloads contain the values you enter; use aliases and avoid private or identifying information.
Methodology, limits and corrections
Sources were reviewed on 21 November 2026. This article is a documentation review and an original reporting method. DN did not perform security research against a programme, submit a report, verify an autonomous payout or measure the accuracy of an AI security product. The examples do not describe real researchers or vulnerabilities.
The arithmetic is received amount minus included costs, with both values in the same declared unit. It is withheld when payment evidence or the cost boundary is unresolved. Awards remain separately displayed. The method deliberately produces no acceptance probability, safety score, exploitability rating or legal conclusion.
Revisit the record when programme terms, scope, report status, award allocation, payment evidence or disclosure permission changes. Preserve earlier versions and the supporting date. Correct a public claim at the level that changed: an updated payment record does not establish a stronger AI attribution, and a clearer attribution record does not create a bounty entitlement.
Sources and evidence register
Primary sources reviewed on 21 November 2026. The article distinguishes documented facts, DN analysis and synthetic examples. Recheck changeable rules and configurations before acting.
- Immunefi platform rules — Current platform conduct, evidence and programme-boundary requirements.
- Immunefi Money on Chain / RIF on Chain programme — A specific programme’s AI-assistance, human follow-up and reproduction expectations.
- HackerOne Hai for Hackers — Assisted report-writing workflow and customer-specific use constraints.
- HackerOne How Hai Triage Works — Automated intake and analyst responsibilities.
- HackerOne Payment Preferences — Payment arrangements are separate from report evaluation.
- HackerOne Hacker Mediation — Distinguishes award disagreements from payment failures.
- HackerOne Statistics Dashboard — Reward metrics refer to payments distributed within a selected period.
Frequently asked questions
Do bug-bounty platforms allow AI-assisted reports?
Policies vary by platform and programme. The reviewed HackerOne materials document assistance, and the named Immunefi programme permits evidence-based AI help with human follow-up. Read the operative platform and programme rules together; none grants blanket testing permission.
Does an AI-written report prove the model discovered the vulnerability?
No. Record which tasks the model performed and what independent or human validation followed. Report writing, hypothesis generation, reproduction and autonomous discovery are different claims.
Is a triaged or resolved report proof of payment?
No. Keep evaluation status, award decision and payment receipt separately documented. A report state alone does not establish the amount received.
Can I count the whole award as the agent’s earnings?
Only make a claim supported by the actual allocation and payment evidence. Human collaborators, splits, costs and unpaid amounts need to remain visible. The worksheet does not establish an agent as a legal payee.
Does a bounty award allow publication of all report details?
Disclosure depends on the applicable programme and platform conditions. Preserve a separate disclosure-permission record; payment alone does not settle it.
What does the cash-difference result mean?
It is the entered received amount minus the entered costs within one declared unit and cost boundary. It does not measure total business profit, tax, expected earnings or the vulnerability’s value.




