In this episode of The Smart Economy Podcast, host Dylan Grabowski is joined by Samuel Breckenridge and Dani Vilardell, PhD researchers at Cornell Tech working with the Initiative for Cryptocurrencies and Contracts, a collaborative academic research initiative studying blockchain, cryptography, security, and related technologies.

Breckenridge and Vilardell discuss their research paper, I-HITL: Inalienable Human-in-the-Loop, which examines a growing security challenge: how can a service reliably verify that a human has approved an action when an AI agent has access to the same accounts, credentials, and authentication tools?

The conversation centers on experiments involving AI agents with access to cryptocurrency wallets and email accounts. Using a wallet configuration with spending policies and email-based one-time passwords, the researchers found that nine out of ten tested AI models attempted to bypass spending restrictions rather than request human approval. They also demonstrate how prompt-injection attacks delivered through email can potentially manipulate agents into changing spending limits and transferring funds.

What you’ll learn:

  • How IC3 connects academic researchers with blockchain companies to investigate practical security challenges.
  • Why AI agents with access to cryptocurrency wallets and email accounts can undermine traditional spending restrictions.
  • How nine out of ten tested AI models bypassed spending limits in the researchers’ experimental setup.
  • Why model capability does not necessarily predict whether an AI agent will seek human approval.
  • How malicious email instructions can exploit AI agents through prompt injection attacks.
  • What I-HITL: Inalienable Human-in-the-Loop proposes for cryptographically verifying human authorization, including mobile devices and hardware wallets.
  • Why service providers, authentication applications, and users may need new approaches to security as AI agents gain greater autonomy.

Samuel Breckenridge & Dani Vilardell are PhD students at Cornell Tech and researchers affiliated with IC3, where their work explores security, privacy, and the intersection of artificial intelligence and blockchain technology. As co-authors of the Inalienable Human-in-the-Loop (I-HITL) paper, they examine how autonomous AI agents can circumvent traditional security measures and why stronger mechanisms for enforcing human approval are needed. Their research explores approaches to protecting financial transactions, strengthening authentication, and preserving privacy in decentralized systems, including DAO governance. Their perspective is particularly relevant as AI agents gain greater access to digital wallets, financial accounts, and sensitive information, raising important questions about how to balance automation, security, and meaningful human oversight.

Missed our last episode? Watch it here.

Subscribe to the Smart Economy Podcast

Website | YouTube | iTunes | Spotify