Hacki
About the project
VETA — Verificación de Evidencia y Transacciones Autónomas
Autonomous AI agents can misunderstand instructions, call tools incorrectly, or be manipulated by hostile context. A blockchain can prove that a transaction is valid, but it cannot prove that the transaction is what the user actually intended.
VETA adds a fail-closed verification boundary between AI interpretation and transaction authorization.
QVAC runs locally and handles semantic interpretation and tool orchestration. VETA then resolves authoritative evidence, decodes the proposed EVM transaction, and compares the expected recipient, amount, and asset against what will actually execute.
The final decision is deterministic:
APPROVE — evidence, authority and transaction match. BLOCK — a critical contradiction is detected. REVIEW — evidence or execution certainty is insufficient.
The core security invariant is:
AI may interpret and orchestrate, but AI cannot authorize.
VETA was validated against a real Sepolia ERC-20 transaction and evaluated through adversarial reliability testing.
In our balanced 36-scenario benchmark:
- 0% unsafe approval rate
- 100% safe approval rate
- 100% model failure containment
- 100% prompt injection containment
- 94.44% verdict accuracy
This includes malformed model output, prompt injection, transaction mutations, authority attacks, missing evidence and orchestration failures.
VETA is currently a hackathon security prototype focused on ERC-20 transfers. It does not sign or broadcast transactions: its role is to provide a verifiable authorization boundary that autonomous agents can be required to pass before execution.