Pearl — Autonomous Bug Bounty Engine
Problem
Bug bounty hunting is time-intensive and manual. Triaging targets, writing payloads, chasing false positives, and escalating findings to maximum impact takes hours per endpoint — and most of that work is repetitive across programs.
Approach
Pearl is a fully autonomous MCP server that plugs into Claude Code. Point it at a scope and walk away.
- Threat-model-first: before touching any endpoint, Pearl correlates the tech stack with historical CVEs and disclosed HackerOne reports to decide which vuln classes to prioritize.
- 8-gate evidence pipeline: every candidate finding must pass exploitation depth, OOB confirmation, semantic diff, endpoint sensitivity, data exfiltration, confidence floor, 3/3 replay, and WAF auto-retry gates. Anything that fails is silently discarded.
- Full escalation chains: SQLi → UNION → DB dump → RCE. SSRF → metadata → IAM creds.
Every chain produces a ceiling, max impact, and working
curlPoC. - OWASP LLM Top 10: prompt injection, jailbreaks, tool hijacking, cross-context injection, and agentic escalation paths.
- Self-learning engine: per-target memory, cross-target patterns, and writeup lessons persist across hunts.
Result
A zero-false-positive hunting engine that ships platform-ready reports (HackerOne, Bugcrowd, Intigriti) with real extracted data, working PoCs, and confirmed escalation ceilings. Currently private.
