AI Receptionist — agentic threat model

Not certain from the listing — The underlying LLM and speech-to-text/text-to-speech models are unspecified. They are potentially vulnerable to voice-based prompt injection (vishing attacks) where callers use adversarial phrasing to bypass safety guardrails or reprogram the agent's behavior during a call.

Not certain from the listing — The agent reads and writes to CRMs and calendars. This creates risks of data exfiltration of customer PII/PHI and database poisoning if malicious caller inputs are written directly into CRM fields without sanitization.

Not certain from the listing — The orchestration framework managing custom scripts, intent detection, and tool execution is proprietary. Vulnerabilities could include insecure tool integration with CRM/calendar APIs, allowing callers to manipulate scheduling logic or access unauthorized records.

Not certain from the listing — The hosting infrastructure and telephony/SIP integration details are not provided. Risks include unauthorized call interception, SIP trunk abuse, or infrastructure compromise leading to access to API keys for integrated CRMs.

Not certain from the listing — It is unclear what logging, guardrails, or drift detection mechanisms are in place. Blind spots in voice-to-text translation could allow malicious payloads to pass undetected to downstream systems.

Not certain from the listing — While industry-specific customization for healthcare is mentioned, there is no explicit confirmation of HIPAA compliance, SOC2 certification, or robust access control policies governing CRM/calendar integrations.

Not certain from the listing — The agent interacts with external ecosystems (CRM and calendar APIs). Cascading failures could occur if these downstream services are compromised, rate-limited, or if the agent is manipulated into flooding them with automated requests.