Mitra — agentic threat model

Not certain from the listing — likely utilizes proprietary or third-party LLMs combined with speech-to-text (STT) and text-to-speech (TTS) models. Key threats include prompt injection via the call recipient's voice input (e.g., instructing the agent to ignore previous instructions) and model reprogramming to bypass safety filters for harassment.

Not certain from the listing — processes and stores highly sensitive conversational data, including call recordings and summaries. Risks include unauthorized access to call logs, data exfiltration of private conversations, and lack of clear data retention policies for voice recordings.

Not certain from the listing — relies on orchestration frameworks to manage call states, handle interruptions, and trigger actions like leaving voicemails. Vulnerabilities include state-machine manipulation where an attacker tricks the agent into executing unintended logic or dialing unauthorized numbers.

Not certain from the listing — requires integration with VoIP/SIP gateways to bridge the AI model with the public switched telephone network (PSTN). Threats include SIP trunk abuse, unauthorized infrastructure access, and lack of sandboxing for telephony session states.

Not certain from the listing — generates call recordings and summaries but lacks visible real-time guardrails to detect if the agent is being used for vishing (voice phishing), spamming, or generating abusive content during live calls.

Not certain from the listing — caller ID spoofing capabilities present severe regulatory and compliance risks (e.g., FCC regulations, TRACED Act, GDPR/CCPA consent requirements for call recording). There is no evidence of robust identity verification to prove the user actually owns the caller ID being used.

Not certain from the listing — primarily designed for human-to-agent interaction over PSTN, but risks include infinite loops or cascading failures if Mitra interacts with other automated voice agents or complex interactive voice response (IVR) systems.