ChainSleuth

Inspiration

ECDSA secures Bitcoin, Ethereum, and most of crypto — unless a wallet reuses the random nonce k when signing. When it does, the private key can be recovered with a few lines of arithmetic. This isn't theoretical: the same flaw broke the PlayStation 3 (2010) and drained Android Bitcoin wallets via the SecureRandom bug (2013). Finding it across thousands of transactions — and proving a compromise without an AI hallucinating a vulnerability — is the problem ChainSleuth solves.

What it does

Point ChainSleuth at a wallet. A team of Qwen-powered agents plans the audit, collects the wallet's signatures, flags any reused signing nonces, and a deterministic verifier recovers the private key and proves it against the on-chain signatures — then writes a forensic incident report with remediation. If a wallet is compromised, ChainSleuth doesn't just claim it — it proves it, then masks the recovered key behind a human-in-the-loop analyst checkpoint (responsible disclosure).

How we built it

Multi-agent pipeline on Qwen Cloud — Planner, Collector, SignatureAnalyst, and Reporter agents run on qwen-max via Alibaba Cloud Model Studio (DashScope), orchestrated around a deterministic CryptoVerifier.
Trust anchor — every cryptographic claim is produced by pure-Python secp256k1 math and verified. The LLM never decides whether a wallet is compromised; that verdict is always mathematical and reproducible.
Real on-chain data — pure-Python keccak-256 + RLP decode raw signed Ethereum transactions (legacy / EIP-2930 / EIP-1559), reconstruct the signing hash, and recover the signer.
Stack — Python + FastAPI backend; React 19 + TypeScript + Vite dashboard with an animated agent pipeline and the analyst key-reveal checkpoint; reports archive to Alibaba Cloud OSS.

Challenges we ran into

Keeping the LLM strictly in an orchestration-and-explanation role. The whole value proposition is that a security finding can never be a hallucination — so the architecture makes it impossible for an agent's narrative to become the verdict. The deterministic verifier is the gatekeeper: no proof, no finding.

Accomplishments that we're proud of

Real ECDSA private-key recovery that verifies, not a mock. A multi-agent system where the AI is genuinely useful (planning, triage, reporting) yet structurally unable to fabricate a result. End-to-end on Qwen Cloud + Alibaba Cloud, with a human-in-the-loop checkpoint modelling responsible disclosure.

What we learned

The most trustworthy way to use an LLM in security tooling is to wrap it around a deterministic core — let it explain and orchestrate, never adjudicate.

What's next for ChainSleuth

Biased / short-nonce (lattice) attacks beyond exact reuse, live multi-chain collection, and batch auditing so exchanges and custodians can sweep thousands of wallets before attackers do.