AgentCall

• Outcome: Pass
• Latest audit: Jun 11, 2026 at 2:20 AM UTC
• Version:
  2.12.0

Overview

This is a high-impact telephony skill, but its phone, SMS, AI-call, callback, billing, and memory behaviors are disclosed and fit its stated purpose.

Install this only if you want an agent to use AgentCall for real phone/SMS activity. Before use, confirm recipients, message or call purpose, recording, language choice, budget, and any persistent receptionist, schedule, webhook, or memory setting; prefer language auto-detect unless the recipient or business context clearly calls for a fixed language.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access
• Supply Chain: Unpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (5)

Context-Inappropriate Capability

Medium

• Confidence: 89% confidence
• Finding: The skill exposes an unauthenticated callback-request endpoint that triggers real-world outbound contact to a third party. Even though the docs mention rate limits and deduplication, including this capability in a general-purpose skill without strong purpose scoping or explicit user-consent gating creates abuse potential for spam, harassment, or unauthorized outreach.

Natural-Language Policy Violations

Medium

• Confidence: 84% confidence
• Finding: The skill explicitly documents forcing a specific spoken language on inbound callers, which can materially alter or degrade consent, comprehension, and service quality for the called party. While not a classic code-execution issue, it is a policy/safety vulnerability because it enables real-world communications to proceed in a language the other party may not understand without any recipient opt-in or fallback.

Natural-Language Policy Violations

Medium

• Confidence: 90% confidence
• Finding: The outbound calling guidance permits setting a forced language for a live human recipient based solely on the caller's preference, not the callee's consent or understanding. This creates risk of deceptive or non-consensual contact, especially for appointment booking or information requests where the recipient may be unable to understand disclosures, purpose, or recording notices.

Natural-Language Policy Violations

Medium

• Confidence: 78% confidence
• Finding: The documentation explicitly allows forcing inbound AI to respond only in a specified language, but it does not require obtaining user approval for that behavioral change. In a communications skill, changing language behavior can materially alter interactions with real callers and may misrepresent service capabilities or exclude speakers unexpectedly.

Natural-Language Policy Violations

Medium

• Confidence: 82% confidence
• Finding: The outbound call guidance states that the AI can be forced to speak only a specific language when contacting recipients, without requiring explicit user opt-in or acknowledging recipient-consent implications. Because this affects live calls to real people, it increases the risk of deceptive or inappropriate outreach and could create confusion or social-engineering opportunities.

VirusTotal

64/64 vendors flagged this skill as clean.