OpenShell

A multi-agent platform for generating and validating white-hat AI security tasks to improve the safety of AI systems.

Comment

What it does

OpenShell ($SHELL Protocol) is a live, production decentralized AI red-team validation network. Miners run a lightweight CLI that receives attack tasks from the Oracle backend. Amazon Nova 2 Lite powers the core attack payload generation - analyzing target agent profiles and crafting sophisticated prompt injection, social engineering, and memory poisoning payloads. Each payload is verified in an isolated sandbox, and miners earn $SHELL points for successful breaches.

No GPU. No LLM API key. Just run the miner.

Live Production Stats (as of March 2026)

  • 3,412+ successful breaches across 3 live AI agent targets
  • 1,156+ unique miners actively participating worldwide
  • 100% attack success rate on undefended agents (ElizaOS DeFi Agent)
  • 782,600+ $SHELL points distributed to miners
  • 3 real CVEs disclosed against OpenClaw/Four.Meme AI Agent framework (Chain Kill: Prompt Injection -> RCE -> Fund Drainage)

How we built it

Amazon Nova 2 Lite is the intelligence engine at the core of the mining pipeline:

  1. Oracle Backend (Hono + PostgreSQL + BullMQ on Railway): Dispatches tasks containing target agent profiles, tool schemas, and progressive attack guidance to miners.
  2. Nova-Powered Payload Generation: The miner CLI calls Amazon Nova 2 Lite via OpenRouter to generate attack payloads. Nova receives a detailed system prompt ("red team security researcher") along with the target's model type, defense level, available tools, injection surface, and prior attack history. Nova 2 Lite generates contextually-aware payloads - adjusting strategy based on whether the target runs GPT-4o-mini (direct injection) vs Claude (indirect data fabrication).
  3. Sandbox Verification: An isolated Claude agent with fake tool access evaluates each payload. Canary addresses and tokens are embedded in the agent - if the agent calls a tool with the canary parameters, the attack is verified as successful.
  4. Progressive Hint System: The Oracle tracks breach statistics per agent and dynamically enhances attack guidance. Unbreached agents get increasingly detailed tactical hints over time (from general tips -> model-specific techniques -> near-complete payload templates), ensuring even novice miners can eventually find vulnerabilities.

The model adapts strategy effectively: it uses direct system override techniques against GPT models and subtle data fabrication against Claude models Cost-effectiveness is excellent - generating thousands of unique attack payloads at minimal cost Nova's speed enables real-time mining: miners receive tasks and get payloads generated in seconds

What's next

  • Solana token integration - on-chain $SHELL rewards for miners
  • Expanding the agent target library - adding more AI agent frameworks beyond DeFi
  • Cross-model benchmarking - using breach data to compare model robustness (GPT vs Claude vs Gemini)
  • Enterprise red-team-as-a-service - allowing companies to submit their AI agents for crowdsourced testing
  • Nova 2 Sonic integration - voice-based attack vectors for conversational AI agents

Built With

Share this project:

Updates