Signal’s Meredith Whittaker DEBATES an AI policy expert: Can we trust AI agents? | Summit 2026

The gist

1. Whittaker says consumer agents need deep root permissions and follow text as commands, making prompt injection a serious way to hijack them.
2. She cites a test where Anthropic's Claude Chrome experiment failed to resist prompt injection 11% of the time, even with guard rails.
3. Chilson argues agents flip the old ad-funded internet bargain, letting ordinary people use their own data for tasks only the wealthy could delegate before.
4. Chilson says he used Muse with a six-word prompt to opt out of five major US data brokers, showing agents can make privacy easier.
   Muse
5. Whittaker counters that Muse leaked addresses, logged contact lists, and was blocked by Chase Bank, and that agents reading screens break Signal's app-level security.