myroSecurity
Human approvals
Myro works freely on routine, reversible tasks — and stops for you on the ones that matter. Outbound and hard-to-undo actions wait for an explicit human decision before they happen.
Gated where it counts
Reading, searching and summarizing flow without friction. But a first email to a new recipient, sharing a file externally, a destructive or irreversible operation — these pause and queue for your approval. Friction is spent only where it earns its place.
First-contact protection
The first time Myro would reach a brand-new destination — a recipient you've never emailed, an unfamiliar external party — it stops and asks. Once you approve a destination, future routine contact flows. New ground always requires a human.
Out-of-band, un-forgeable
Approval cards are rendered by the control plane from its own record of the action — the exact recipient, subject and content — never from text the agent generated. The agent cannot dress up a request, cannot approve on your behalf, and cannot see or influence the approval channel.
Unattended work stays bounded
Scheduled and overnight tasks obey the same rules. Anything that would require approval queues for you instead of executing. An agent working while you sleep can never do more than one working under your eye.
Why it matters
Prompt-injection and over-eager automation are real: a crafted email or web page can try to steer an agent into sending data somewhere it shouldn't. First-contact approvals mean that even if Myro is misled, it cannot open a new outbound channel without a human tap. You hold the irreversible decisions.