Goodcall AI — agentic threat model

Not certain from the listing — The specific underlying LLM or speech-to-text/text-to-speech models are not disclosed. The primary threat is voice-based prompt injection (vishing) where callers manipulate the model into bypassing business logic or leaking system prompts.

The agent utilizes custom skills training and lead capture, implying a local knowledge base or dynamic prompt construction. Threats include data poisoning of the custom skills training set and exfiltration of captured customer lead data or CRM-stored PII.

The agent orchestrates call flows, appointment booking, and CRM writes. Insecure tool integration is a major threat, where malicious callers could exploit tool execution paths to manipulate CRM records or book fraudulent appointments.

Not certain from the listing — The hosting environment, API gateway security, and sandboxing of the integration layer are unspecified. Threats include exposed webhook endpoints and insecure storage of CRM API keys.

Not certain from the listing — There is no mention of real-time guardrails, call transcript monitoring, or anomaly detection for unusual call volumes or malicious injection patterns.

Not certain from the listing — Compliance certifications (such as SOC2, HIPAA, or PCI-DSS) are not stated, which is a concern given that the agent handles customer PII and schedules appointments.

The agent integrates directly into external business workflows and CRMs. The primary threat is downstream cascading failures, where a compromised agent acts as a trusted entity to inject malicious data into third-party CRM ecosystems.