Masters of Scale Summit

Can we trust AI agents with our data? A Signal debate

Key points

Key takeaways from the Masters of Scale Summit live debate between Signal president Meredith Whittaker and Abundance Institute head of AI policy Neil Chilson on whether we should trust AI agents with our information (November 2026):

The resolution. Should we stop trusting AI agents with our information? Whittaker argued we should not trust fallible systems, while building trustworthy ones.

Root permissions. To book a birthday party an agent needs your contacts, credit card, calendar and messages — breaking down the sandboxes that are a core paradigm of security engineering.

Prompt injection is the fatal flaw. A zero-click exploit pulled corporate data off Microsoft Copilot; Perplexity's agentic browser was hacked via a Reddit comment; and 11% of the time Anthropic's Claude Chrome experiment could not resist prompt injection even with guardrails.

Meta's Muse under fire. Whittaker cited Muse logging people's contact lists, publishing private company data, and sending a buyer to a stranger's home; Chase Bank blocked Muse over authentication and liability concerns.

Agent architecture resembles malware. Pervasive permissions, screen capture, off-site inference and driving your browser as if it were you — the same pattern as targeted malware.

Chilson's counter. Agents flip the internet's data-for-ads bargain so data is used for your own purposes; he used Muse to unregister himself from five data brokers with a six-word prompt.

Verification, not trust. Whittaker's standard: 'trust comes with verification' — Signal is verifiably secure because anyone can inspect its code.

The genie problem. A consumer agent that does everything for you needs to know everything about you; one that checks in at every step is not really an agent.

Read more: Masters of Scale

More top stories