autohack

An autonomous security researcher that hunts real bug-bounty programs end to end.

autohack discovers bounty programs across five platforms, spawns hour-long Claude sessions to hunt for vulnerabilities, validates its own findings through adversarial self-review, drafts platform-ready reports, and submits them — with a human approval gate in front of anything that goes out. It learns across hunts: every outcome, near-miss, and triager response is written to a memory store so the next session starts with context from every past one.

_{The real-time tRPC dashboard. Top: the tiered-model control (Scout / Hunt / Review / Submit, each assignable to Opus, Sonnet, or Haiku) and the live pipeline. Bottom: findings moving through the lifecycle with per-finding confidence and status. Shown with sample data.}

Built with direct Anthropic SDK calls — no LangChain, no CrewAI, no framework wrappers.

How it works

1. Discovery — polls five bounty platforms (HackerOne, Immunefi, Huntr, plus an aggregator covering Bugcrowd, Intigriti, and YesWeHack), deduplicates targets, and ranks them by payout, scope, freshness, and saturation.
2. Analysis — pulls each target's scope, prior reports, and asset inventory, then produces a hunt strategy and feasibility score.
3. Hunt — spawns Claude in a pseudo-terminal for a 60-minute session with a scoped cheat sheet. It works autonomously on recon, exploit discovery, and validation, streaming every tool call to the dashboard in real time.
4. Adversarial self-review — a second Claude instance attacks the finding from the opposite direction, scores it on a 0–15 binary rubric, and tries to disprove it. Anything below the threshold is rejected before it becomes a draft.
5. Report drafting — generates a platform-formatted report with reproduction steps, impact analysis, and remediation. A separate Sonnet pass compresses verbose findings before submission.
6. Submission — optionally auto-submits through the platform API, gated behind a manual approval flag.

Findings move through a 12-state lifecycle, visible on the board above:

discovered → scanning → analyzing → validated → drafting → reviewing
   → submitted → triaged → accepted / rewarded / duplicate / rejected

Highlights

• Adversarial validation gate. A finding only advances from scanning to drafting after a separate Claude instance tries to refute it and scores it ≥ 8/15. This filters hallucinated findings before they ever reach a report.
• Tiered models per stage. Scout, Hunt, Review, and Submit are each assignable to Opus, Sonnet, or Haiku with independent effort levels — spend Opus where it matters, Haiku where it doesn't. Configurable live from the dashboard.
• Cross-hunt memory. Every hunt writes findings, near-misses, and dead ends to a persistent store; new hunts are primed with relevant prior context so the model doesn't re-explore known dead ends.
• PTY-based Claude spawning. Hunts run in a pseudo-terminal so output streams live to the dashboard via xterm.js, with a hard 60-minute timeout and graceful shutdown.
• Ephemeral prompt caching cuts input tokens ~90% across repeated hunt sessions. Dual backends: Claude Max via the CLI, or the API via SDK (CLAUDE_BACKEND=cli|api).
• Crash-safe coordination. A cross-process lock file plus stale-PID detection stops two hunts from touching the same target; error classification (transient / permanent / validation / timeout) decides whether to retry, skip, or kill.

The same harness also runs a bounty agent on the Algora platform: it spawns Claude Code sessions for long autonomous runs, executes the test suite, opens PRs, and addresses review feedback on its own.

Architecture

packages/
  core/                 shared config, DB (SQLite + Drizzle), logging, Claude wrappers
  security-discovery/   HackerOne, Immunefi, Huntr, Bugcrowd, Intigriti pollers
  security-analyzer/    target ranking and feasibility
  security-solver/      PTY-based Claude spawning, 60-min hunt sessions
  security-memory/      cross-hunt outcome and near-miss learning
  dashboard/            Next.js 15 + tRPC real-time monitoring UI
orchestrator.ts         cron-scheduled hunt loop

Stack

TypeScript · Node.js · Next.js 15 · tRPC · SQLite + Drizzle ORM · Anthropic SDK · xterm.js · Pino · node-cron · pnpm monorepo.

Running locally

pnpm install
cp .env.example .env    # set HACKERONE_API_TOKEN, IMMUNEFI_API_KEY, etc.
pnpm dev                # orchestrator + dashboard on :3456
pnpm dashboard          # dashboard only (no hunts), on :3456

Scope and ethics

autohack only hunts targets with active, in-scope bug-bounty programs, and never attacks systems without authorization. The cheat sheet and prompt scaffolding explicitly bound Claude to each program's declared scope. Every submission passes through a manual approval gate unless auto-submit is explicitly enabled.