# I Built an AI Bug Bounty Hunter. Here's What It Actually Found.

- **Category:** AI
- **Date:** 2026-02-24
- **Read time:** 8 min

Most AI demos are impressive until you ask "does it actually work?" I built an autonomous bug bounty hunting system: one command, six specialist AI agents, real vulnerabilities found on real programs. Then I open-sourced it.

## The architecture

A single Claude Opus orchestrator delegates to six specialist agents, each with restricted tool access: Scope Analyzer, Recon Agent, Web Vuln Agent, API Vuln Agent, Source Review Agent, and Report Writer. Built on the Claude Agent SDK with 19 custom MCP tools across 6 servers.

## What it found

On the first successful run against a real HackerOne program: 30 findings (7 medium, 16 low, 7 informational). 22 minutes, 36 turns, $8.91 in API cost. We're not naming the program or specific endpoints. Findings were reported through proper disclosure and some are still in triage. Categories included exposed source maps, leaking configuration endpoints, XML-RPC vulnerabilities, and exposed CA keys, each with HTTP request/response evidence.

## Key technical decisions

- Orchestrator-worker pattern: Opus for strategy, Sonnet for execution
- Tool restrictions enforced at SDK level, not by prompting
- Incremental persistence: findings written to disk immediately
- Evidence validation: no finding stored without concrete proof
- AsyncIterable prompt fix to keep MCP transport alive
- Response truncation at 500KB to stay under SDK buffer limits

## Why open source

The architecture patterns (orchestrator-worker with MCP tool restrictions, evidence validation, transport fixes) apply well beyond bug bounty. The full source is at github.com/eliBenven/aibugbounty (MIT licensed).