CallFluent AI — agentic threat model

Not certain from the listing — The underlying LLM and speech synthesis/recognition models are unspecified. Key threats include voice-based prompt injection (VUI injection) to bypass behavioral guardrails, and adversarial audio inputs designed to manipulate the agent's logic.

Not certain from the listing — The agent interacts with CRM data to personalize calls. Threats include unauthorized exfiltration of customer PII during voice sessions, CRM data poisoning, and insecure storage of sensitive call transcripts or audio recordings.

Not certain from the listing — The orchestration layer manages the state machine of the call and CRM tool execution. Threats include insecure tool integration where a manipulated conversation triggers unauthorized CRM API writes or data deletion.

Not certain from the listing — The hosting environment must secure telephony (SIP/VoIP) infrastructure and API secrets. Threats include SIP trunk hijacking, toll fraud, and unauthorized access to API keys used for CRM integrations.

Not certain from the listing — It is unclear what real-time guardrails or monitoring exist. Threats include conversational drift, lack of real-time injection detection in the audio stream, and insufficient logging of anomalous agent behavior.

Not certain from the listing — The platform must adhere to strict telephony regulations (e.g., TCPA, GDPR for voice recording consent, PCI-DSS if payments are discussed). Threats include non-compliance due to lack of explicit recording consent mechanisms or insecure handling of spoken credentials.

Not certain from the listing — The agent operates primarily as a point-to-point integration with CRMs rather than a multi-agent network. Threats are limited to cascading trust failures between the voice agent and the connected CRM platform.