ClawHub
Back to skill

Security audit

Talking Avatar

Security checks for vulnerabilities and agentic risk

Overview

The skill is coherent and non-executable, but it enables realistic person-and-voice video generation without clear consent or anti-impersonation safeguards.

Install only if you are comfortable enforcing your own policy checks before use: confirm the user owns or is authorized to use the portrait, footage, and voice; avoid impersonation or deceptive re-voicing of real people; and disclose synthetic or altered media where appropriate.

SkillSpector

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (1)

Missing User Warnings

Medium
Confidence
94% confidence
Finding
This skill explicitly enables re-voicing footage and making a portrait or avatar speak, but it provides no consent, authorization, or anti-impersonation guardrails. In this context, the omission materially increases the risk of non-consensual deepfakes, deceptive impersonation, and reputational or fraud harms because the workflow normalizes powerful face+voice synthesis without any safety gating.

Static analysis

No suspicious patterns detected.