Signal President Meredith Whittaker and VP Udbhav Tiwari warned that agentic AI is insecure, unreliable, and a surveillance nightmare. Speaking at Hamburg’s 39C3 conference, they argued that OS-level integration, like Microsoft’s Recall feature, creates a single database of user activity vulnerable to malware and prompt injection attacks. These systems bypass end-to-end encryption, turning personal devices into forensic dossiers accessible to attackers.

Reliability is another core problem. Whittaker explained that AI agents are probabilistic, meaning each step in a complex task degrades accuracy. Even with high per-step success rates, multi-step actions quickly fail, with current models falling short roughly 70% of the time. This unreliability, combined with deep data access, creates a dangerous mix for enterprise and consumer trust.

The pair demanded immediate industry restraint: stop reckless deployment, make privacy the default, and force radical transparency. Without these guardrails, the nascent agentic AI push risks a massive trust collapse. Consumers are already wary of overhyped tech, and security failures could stall investment. The message is clear: fix the architecture before it scales.