MinutesLink — agentic threat model

Not certain from the listing — likely utilizes third-party speech-to-text and LLMs for transcription, summarization, and avatar generation. Key threats include prompt injection via meeting audio/chat and model reprogramming during avatar interaction.

High risk. The agent ingests, processes, and stores highly sensitive meeting audio, transcripts, and user interaction history over time to train personalized avatars. This creates a high-value target for data exfiltration, unauthorized access, and training data poisoning.

The agent orchestrates calendar monitoring, autonomous meeting joining, audio recording, and avatar task execution. Vulnerabilities include insecure tool integration with Google Workspace APIs and unauthorized task execution by the trained avatars.

Not certain from the listing — requires infrastructure to host the bot that joins calls, process real-time audio streams, and store recordings. Threats include container compromise, unauthorized access to cloud storage buckets containing raw audio, and insecure API endpoints.

Not certain from the listing — requires robust guardrails and observability to monitor avatar conversations and prevent them from generating harmful, hallucinated, or unauthorized statements when representing the user.

Critical compliance and security layer. The agent must handle OAuth permissions for Google Calendar/Meet securely, comply with regional wiretapping/recording consent laws, and enforce strict access controls to prevent unauthorized sharing of meeting minutes.

The deployment of personalized avatars to 'engage in conversations' and 'monetize expertise' introduces severe identity and trust risks, including avatar hijacking, social engineering of external parties, and unauthorized representation of the user.