Report Bugs From Your AI Agent

Write up, submit and follow your bug reports from Claude Code, Codex or Cursor.

What It's For

You found the bug and reproduced it. Your agent already has your notes, requests and proof of concept in front of it, so it can write the report, check it against the program's scope and submit it through the REST API.

It's for writing up real findings faster. It won't turn a guess into a bug: Bugbop's AI scope check turns away reports with no evidence the bug was reproduced, and pushing them through anyway earns strikes.

Setup

Install the Bugbop skill. In Claude Code:

/plugin marketplace add PonnyPtyLtd/bugbop-skills
/plugin install bugbop@bugbop-skills

For Codex, Cursor and anything else that reads Agent Skills, copy skills/bugbop from the repo into your tool's skills folder. It needs python3 and nothing else.

Then create a token on the API Tokens tab of your profile with the read and write scopes, and export it:

export BUGBOP_TOKEN=bugbop_pat_...

The token is shown once and acts as you, so treat it like a password. Don't paste it into the chat.

Submitting a Report

Point your agent at what you have and ask it to write the report. It shows you the report before anything is sent.

> I found an IDOR on Acme's invoices API. My notes and requests are in notes/acme-idor.md. Write it up and submit it.

⏺ Bash(bugbop.py GET /programs/1a2b3c4d)

⎿ Acme: app.acme.example and api.acme.example in scope, self-XSS out of scope

⏺ Read(notes/acme-idor.md)

⏺ IDOR in /api/invoices/:id returns other accounts' invoices

Steps: log in as account A, request /api/invoices/8812 (owned by account B), get a 200 with B's invoice.

Impact: any signed-in user can read every customer's invoices by counting up the id.

Both requests and responses from your notes are included. api.acme.example is in scope. Submit it?

> yes

⏺ Bash(bugbop.py POST /programs/1a2b3c4d/reports)

⎿ 202, being checked against the scope

⏺ Bash(bugbop.py GET /report_submissions/c0ffee00)

⎿ completed

⏺ Submitted, and the AI rated it high. Add your screenshots on the report page.

The session above is an example, not real output. A good report still needs what the Bug Reporting Guide asks for: steps anyone can follow, the impact and proof.

Scope Check and Strikes

On programs with AI triage, a report is checked against the program's scope before it's saved. That usually takes 10-30 seconds, and you can only have one report being checked at a time, website submissions included.

If the AI thinks the report is out of scope, the skill stops, shows you why and asks what to do. You can fix the report, drop it, or tell it to submit anyway. A report submitted anyway is flagged for the program's team. If they close it as Not Applicable you get a strike, and 3 strikes within 30 days suspends you from submitting reports to any program for a month.

Your agent doesn't get a different rule. A report it submits is your report.

Following Up

The same token reads your reports and their comments, so you can ask:

  • "Which of my reports are waiting on me?"
  • "What did Acme ask on the CSRF report? Draft a reply from my notes."
  • "Add the second request to the invoices report."

Comments are shown to you before they're sent. They notify the program's team and can't be unsent.

Limits

  • The API takes Markdown text. Screenshots, video and HAR files are added in the browser: open the report and edit it, or attach them to a comment.
  • New reports, edits and comments share a limit of 20 an hour and 3 a minute, across all your tokens.
  • On programs with AI triage, the AI sets the severity. Any severity you send is ignored.
  • You can submit to public programs, and to private or Restricted ones you've been let into. Asking to join a Restricted program is done on the program's page.
  • Payouts aren't in the API. Request them from your dashboard.

Find a program to test

Browse public programs and start submitting reports.