Mindy AI — agentic threat model

Not certain from the listing — likely relies on third-party foundation models (e.g., OpenAI, Anthropic). The primary threat is indirect prompt injection, where malicious instructions embedded in incoming emails hijack the model's behavior.

Not certain from the listing — likely processes and indexes inbox messages, calendar events, and search results. Threats include data exfiltration of sensitive email contents and poisoning of the agent's context window with malicious email data.

Not certain from the listing — orchestrates tools for email drafting, scheduling, and deep search. Insecure tool integration could allow an attacker to trigger unauthorized email sends or calendar modifications via manipulated agent planning.

Not certain from the listing — operates as a cloud-hosted SaaS. Requires secure storage of OAuth tokens for email/calendar access; compromise of the hosting infrastructure could expose access tokens for all user inboxes.

Not certain from the listing — no public details on guardrails or output monitoring. Without robust input/output filtering, the agent may silently execute malicious instructions received via email without alerting the user.

Not certain from the listing — requires highly permissive OAuth scopes (read/write/send email, manage calendar). The lack of explicit security certifications or detailed privacy policies in the listing increases compliance and data privacy risks.

Not certain from the listing — primarily functions as a personal assistant, but interacts with external email ecosystems. It faces trust-boundary threats when interacting with other automated email responders or scheduling agents.