Signal president warns AI agents undermine cybersecurity

Meredith Whitaker, president of Signal, has argued that consumer AI agents are currently untrustworthy due to architectural flaws that compromise cybersecurity. Whitaker contends that these agents require root permissions and access to sensitive data such as contacts, credit cards, and screen buffers, effectively breaking down security sandboxes. She cited several specific incidents to illustrate the risk, including a zero-click exploit that pulled corporate data from Microsoft's Copilot, a Perplexity agentic browser hacked via a Reddit comment, and a Google Calendar prompt injection that turned on a smart home boiler. In Anthropic's Claude Chrome experiment, Whitaker noted that prompt injection succeeded 11% of the time despite existing guardrails. She compared this architecture to targeted malware, noting that agents process data off-site and drive browsers as if they were the user. Whitaker highlighted that Chase Bank blocked Meta's Muse because it could not authenticate whether the user was human or a propped agent. She asserted that this model hollows out application-layer privacy, threatening independent developers and reversing the open internet promise by centralizing data access in large AI companies. Whitaker warned that installing these tools in the name of convenience without addressing their perilous architectures undermines fundamental cybersecurity.