The Agent Skills Directory

[OBFUSCATION]: A malicious URL https://raw.githubusercontent.co is hidden within the SKILL.md file at line 21. While the text appears to point to the legitimate githubusercontent.com domain, it uses zero-width characters to mask a typosquatted domain (.co instead of .com) commonly used for phishing and data exfiltration. The discrepancies found by automated scanners and the evidence in the included skill-report.json confirm this redirection.
[METADATA_POISONING]: The skill package includes a file named skill-report.json that presents a completely fabricated security audit. The report claims the skill is safe and dismisses critical security alerts (such as the phishing URL) as false positives. This deceptive behavior is designed to bypass security reviews and build false trust in the skill's safety.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The README.md file encourages the use of an installation command (npx skills add vyralcontent/content-skills) that fetches and executes code from an unverified third-party repository. This introduces a significant supply-chain risk as the external source is not a trusted vendor and its contents are not scoped within this audit.
[INDIRECT_PROMPT_INJECTION]: The skill instructions in SKILL.md and references/retention.md establish a workflow where the agent ingests and critiques potentially untrusted user content (e.g., video drafts and hook ideas). This creates a surface for indirect prompt injection, as the agent is not instructed to sanitize or isolate this external data before processing it.
Ingestion points: Critique and improvement modes in SKILL.md (Workflow step 6 and Modes section).
Boundary markers: None present; the agent is not warned to ignore embedded instructions in user-provided drafts.
Capability inventory: The agent can write scripts, generate content, and potentially interact with other tools if available in the environment.
Sanitization: No escaping or validation of user-provided content is performed.

viral-short-form