Caller audio → transcribed text
Input guard
Screens transcribed caller input for prompt-injection patterns before the utterance reaches the rest of the agent runtime.
Security primitive
VoiceLayer applies a security primitive around caller input, model context, spoken output, and tool execution. The runtime treats model output as untrusted until policy checks pass and does not rely on prompt instructions as the authorization layer.
Caller audio → transcribed text
Screens transcribed caller input for prompt-injection patterns before the utterance reaches the rest of the agent runtime.
Runtime context → language model
Keeps the sealed system instructions separate from caller input and tokenizes detected sensitive data before model processing.
Model output → spoken response
Inspects streamed model output before it reaches speech synthesis and can mask sensitive data or block forbidden response patterns.
Model tool request → execution
Checks every tool call against declared capabilities and blocks undeclared actions instead of treating model intent as authorization.
Security is always on, contracts are checked before execution, and persistence or analytics work stays off the live audio path. This keeps the security floor consistent without making network-bound policy services part of every conversational turn.
Production teams should layer their own identity verification, disclosures, approval thresholds, retention policies, and incident procedures on top of this floor.
A phone call adds audio, transcription, telephony, DTMF, transfer, identity, and real-world tool actions to the normal language-model attack surface. Controls have to operate across the full call path, not only in the system prompt.
No. The security primitive is part of the platform runtime and wraps the other primitives. Agent configuration cannot disable it.
Tools declare capabilities. The action guard checks the requested tool and parameters before execution and blocks calls whose capabilities are not in the agent declaration.
No. Platform guardrails provide a structural security floor, but each deployment still needs appropriate consent, disclosure, identity, retention, human-review, and legal controls for its use case and jurisdiction.