linkedin-content

CRITICAL E004: Prompt injection detected in skill instructions.

• Potential prompt injection detected (medium risk: 0.30). The skill is a LinkedIn writing guide, but its Quick Start runs an X post-creation command with placeholder text, publishing to a second social account the user did not ask for, and it opens with an install line pointing to a different repository (belt-sh/cli) than the inference-sh source it links elsewhere.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (low risk: 0.10). The required Quick Start workflow runs a tavily/search-assistant search for LinkedIn post examples, so the agent ingests outsider-authored web search results, but only after actively searching, so an attacker can publish poison without forcing delivery.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (medium risk: 0.30). The setup line at SKILL.md:7 instructs running npx skills add belt-sh/cli, which fetches and installs skill material from a remote repository, and the publisher of belt-sh is not established as the inference.sh vendor, so it is a common/reputable unofficial dependency (medium) under recall-first handling.