AI bug bounty report writing
AI bug bounty report writing works when the model drafts from your verified notes and you check every line before you submit. Give it the finding, your steps and your evidence, and ask for the platform's sections.
What each platform asks for
From each platform's own report guide.
| Platform | Sections | Severity | On reproduction, quoted |
|---|---|---|---|
| HackerOne | Title, steps to reproduce, expected vs. actual behavior, impact, supporting material | CVSS 3.0, 3.1 or 4.0 | "Provide clear, numbered steps. Include necessary details like URLs, affected parameters, and user roles required to trigger the vulnerability." |
| Bugcrowd | Summary title, target, technical severity, URL or location, description (overview, walkthrough and PoC, evidence, demonstrated impact), attachments | VRT | "Add clear and descriptive replication steps so that the organization can easily reproduce and validate your findings." |
| Intigriti | Title, asset, endpoint, vulnerability type, severity, reproduction details and PoC, impact, recommended solution | CVSS calculator or manual selector | "All information necessary to reproduce the vulnerability should be written in the report, images can be used as guidance, but should not be required to understand and reproduce the submission." |
| YesWeHack | Description (CWE), discovery, proof of concept, exploitation, impact, remediation, references | Not in the guide | "These steps must be easy to follow so that triagers can reproduce your exploit and validate your report as quickly as possible." |
Five steps
- Write raw notes first
Asset, what you did, what happened, what you expected, evidence files. Redact tokens and other users' data.
- Ask for a draft in the platform's sections
Use the prompt here, or the report template.
- Re-run every step from the draft
If a step fails, fix the step, then the wording.
- Write the impact yourself
Then let the model tighten it. See impact statements.
- Disclose AI use where asked
Intigriti requires it, and so do curl and Django.
The prompt
Draft a bug bounty report for HackerOne from my notes below.Sections: Title, Summary, Steps to reproduce, Expected vs. actual behavior, Impact, Supporting material.Rules:- Use only facts in my notes. Add no steps, endpoints, versions or impact.- Mark anything you are unsure of with [CHECK].- Number the steps and name the account role each one needs.- Keep it under 400 words. <notes>[asset, what you did, what happened, what you expected, evidence file names]</notes>What AI does well, and what stays yours
| Part | The model | You |
|---|---|---|
| Title | Tightens the wording | Names the right asset and bug class |
| Steps | Orders and numbers them | Re-runs each one |
| Impact | Removes vague chains of "could lead to" | Decides what you proved |
| Severity | Explains each CVSS metric | Picks the score |
| Evidence | Describes screenshots and recordings | Chooses and redacts attachments |
Titles, before and after
| Before | After |
|---|---|
| XSS on site | Stored XSS in profile bio on app.example.com runs script in any viewer's session |
| IDOR in API | Any signed-in user can read other users' invoices on api.example.com by changing the invoice ID |
| Info leak | Unauthenticated /debug page on api.example.com returns internal service names |
What gets reports closed
"Reports that appear to be unverified AI output, contain fabricated elements, or demonstrate a lack of technical understanding may be closed without response, subject to longer validation times and could lead to removal from the platform."
Intigriti Community Code of Conduct, read November 18, 2026
"Avoid lengthy descriptions generated by AI tools."
Apple Security Bounty guidelines, read November 18, 2026
Models for reports
| Model | Why | In / out per 1M |
|---|---|---|
| Kimi K3.1 Pro | Reads your screenshots and screen recordings | $3.50 / $8.00 |
| DeepSeek V4.1 Flash 1017 | Cheap first drafts and rewrites | $0.24 / $0.90 |
Drafts hold unpatched findings, so where the text goes matters. Silk Compute does not store your prompts or model outputs. See its models.
Questions
Will triage reject a report because AI helped write it?
Not on these four platforms, which allow AI. They reject reports that are unverified, fabricated or not reproducible.
Should the model pick the severity?
Let it explain the metrics, then score it yourself in the platform's scale. HackerOne uses CVSS and Bugcrowd uses the VRT.
How long should a report be?
As long as the steps need and no longer. Apple asks hunters to avoid lengthy AI-generated descriptions.