Batai Assistant — agentic threat model

Not certain from the listing — likely utilizes proprietary or fine-tuned LLMs combined with speech-to-text, text-to-speech, and emotion recognition models. Key threats include adversarial voice inputs, prompt injection via spoken audio, and model reprogramming.

Not certain from the listing — likely processes voice recordings, transcripts, and customer profiles for personalization. Key threats include data exfiltration of sensitive PII from voice/email logs and lack of data lineage for RAG or personalization systems.

Utilizes function calling and third-party integrations to orchestrate actions across phone, WhatsApp, and email. Key threats include insecure tool integration, tool misuse, and injection attacks via incoming messages/emails triggering malicious function calls.

Not certain from the listing — likely hosted on cloud infrastructure to support low-latency optimization and scalability. Key threats include exposed API endpoints, lack of sandboxing for third-party integrations, and infrastructure compromise.

Not certain from the listing — requires low-latency monitoring and emotion-tracking evaluation. Key threats include blind spots in detecting prompt injections over voice channels and insufficient logging of dynamic function calls.

Not certain from the listing — must comply with telecom regulations and data privacy laws (GDPR/CCPA) regarding voice recordings and emotion data, but no specific compliance certifications or security controls are mentioned.

Not certain from the listing — primarily operates as a single-agent platform integrated into business channels, though third-party integrations could introduce cascading failures if external APIs are compromised.