Signal's Whittaker warns AI agents break device security

"We are eating the seed corn of our cybersecurity when we install these things in the name of convenience without thinking through the perilous architectures." That was Meredith Whittaker, president of Signal, arguing against trusting AI agents in an Open to Debate event staged at the Masters of Scale Summit in San Francisco.

Her opponent, Neil Chilson, head of AI policy at the Abundance Institute, pitched agents as a privacy tool rather than a privacy problem. "Being able to file and register and opt out of all the data brokers with a six-word prompt is pretty powerful privacy protection," he said. His case was comparative: "The status quo is not great."

Whittaker's technical argument ran through worked examples. Meta's Muse, she said, "is logging people's contact lists, logging the people they talk to...creating dossiers about me and your communications." The architecture itself is the break: "If an agent is screenshotting everything on your device...you have just broken Signal's security model." She cited an Anthropic Claude/Chrome experiment in which, per her telling, "11% of the time, even with guardrails, it was not able to resist prompt injection."

Chilson compared the current state to the 1990s virus problem: not solved, but managed. Jeff Berman, CEO of WaitWhat, moderated. The full debate transcript is published by Open to Debate.