Saner.AI — agentic threat model

Not certain from the listing — likely relies on third-party commercial LLMs. The primary threat is indirect prompt injection via incoming emails, which could manipulate the model into executing unauthorized actions or leaking sensitive notes.

Not certain from the listing — likely utilizes a vector database to perform RAG over the user's personal notes and emails. Threats include data exfiltration of sensitive personal history and potential knowledge-base poisoning from malicious incoming emails.

Not certain from the listing — orchestrates note-taking, task management, and email APIs. Insecure tool integration could allow an attacker to trigger unintended email dispatches or delete critical tasks through manipulated context.

Not certain from the listing — hosted as a closed-source SaaS platform. The critical threat at this layer is the secure storage and handling of third-party OAuth tokens (e.g., Google/Microsoft) used to access user email and calendar accounts.

Not certain from the listing — no public details on guardrails or output monitoring. A lack of input/output filtering could allow prompt injections to succeed silently, leading to data leakage or incorrect task generation.

Not certain from the listing — requires robust identity and access management to isolate user data. Compliance with privacy regulations (like GDPR) is critical given the highly personal nature of ADHD-focused note-taking and email processing.

Not certain from the listing — operates primarily as a single-user personal assistant. However, interaction with external email ecosystems introduces risks of cascading trust issues when processing untrusted external communications.