The Draft Approval Gate: Making AI Email Agents Safe by Default

As AI agents increasingly handle sensitive communications like email, relying on prompts alone for safety is dangerously insufficient—prompt injection attacks can bypass guardrails and send unauthorized messages. The solution is elegantly simple: split send permissions from write permissions at the infrastructure level, forcing all emails through human-approved drafts where attackers cannot reach. Nylas Agent Accounts make implementing this secure-by-default pattern straightforward, transforming safety from a soft organizational rule into an impossible-to-bypass architectural guarantee.