AI agents are capable and easily fooled at the same time. They can read your data, act on your behalf, and reach the outside world. Because LLMs don't fundamentally separate data from instructions, anything they read can be treated as a command.
In this episode of The Ledger Podcast, Leonard Lin, CTO of Shisa.ai, in conversation with Ledger Chief Human Agency Officer Ian Rogers, talks about building AI agents you can trust. Leonard's framework, Shisa D, inverts the usual control plane so the AI proposes and the system decides. Secrets are removed from the model, tool calls are reviewed by a separate layer, and the highest-risk actions route to a Ledger signer, where you see the real intent in plain text and approve it with a physical tap.
Plus a live demo of Shisa D routing an action to a Ledger signer for hardware approval.
","author_name":"Ledger"}