Security audit

Azure Voicelive

Security checks across malware telemetry and agentic risk

Overview

The skill is a plausible Azure voice-integration guide, but it needs review because it handles caller/customer data and credentials while containing confusing, mismatched documentation.

Review this before installing in any real customer-service or telephony environment. Treat it as a draft Azure VoiceLive integration guide: clarify which credentials are required, avoid putting full caller identifiers or order data into prompts, document that data is sent to Azure, and remove the unrelated grading output schema.

SkillSpector

By NVIDIA

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (5)

Intent-Code Divergence

High

Confidence: 92% confidence
Finding: The documented output schema describes a grading/audit report unrelated to the actual Azure real-time voice functionality. This mismatch can cause an agent or user to mis-handle outputs, trust incorrect semantics, or route sensitive data through the wrong workflow, which is a security-relevant integrity problem even if it is not direct code execution.

Intent-Code Divergence

Medium

Confidence: 89% confidence
Finding: The skill documents API key setup in a way that conflicts with the code samples, which use Azure Cognitive Services endpoint/key variables and sometimes managed identity. Contradictory credential guidance can lead operators to expose, mis-store, or misuse secrets, and may cause fallback to weaker authentication practices.

Missing User Warnings

Medium

Confidence: 91% confidence
Finding: The skill processes caller and order-related data through an external cloud AI service but does not clearly warn users that potentially sensitive personal and transactional information is being transmitted off-host. In a telephony/customer-service context, this increases privacy and compliance risk because users or operators may not realize what data leaves their environment.

Missing User Warnings

Low

Confidence: 80% confidence
Finding: The documentation references API keys and sensitive environment variables without sufficiently surfacing credential-handling risks in the overview. While examples commonly use environment variables, the absence of explicit secure-handling guidance can contribute to accidental leakage through logs, shell history, screenshots, or source control.

Ssd 3

Medium

Confidence: 94% confidence
Finding: The example injects caller phone information directly into the model conversation context as natural-language system content. This creates a straightforward path for the model to reveal or repeat private information in responses, especially in voice interactions where disclosures may be hard to monitor or retract.

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.