who'se Agent Is It Anyway - Agency, Authorization, and Accountability

The talk starts from a simple question: when someone asks an AI agent to shop on Amazon, who has approved that agent's access? Websites have always granted access to people, and tools have never needed their own permission. The speakers lay out six competing interests, including user choice, user security, platform costs and trust, and competition, before turning to the Amazon v. Perplexity litigation under the Computer Fraud and Abuse Act.

They explain how the CFAA is read broadly, reaching automated access and carrying both criminal and civil exposure. The district court sided with Amazon on authorization, but the Ninth Circuit reversed on the access element, reasoning that the agent functions as a tool and the user is the one reaching the servers. The talk notes that this leaves other agent designs open and that other circuits may rule differently. It also covers Amazon's own shopping agent, commerce protocols, and the concern that agent access is settled through B2B deals rather than by users.

The second half looks at a discussion draft of a federal AI Agent Act. It would let users designate a registered custodial agent, with the FTC gatekeeping registration and NIST setting technical protocols, plus cost-shifting rules for platforms. The speakers argue the draft does not handle disputes where a user never revoked an agent that then made an unwanted purchase. They close by pointing to open comment periods as a way for practitioners to shape the rules.