The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The file skill-report.json contains self-referential instructions and metadata declaring the skill as "safe" and preemptively dismissing potential security findings as false positives. This technique attempts to influence the auditor's judgment and automated analysis tools.
Ingestion points: The entire content of skill-report.json acts as an untrusted data source within the skill package.
Boundary markers: None present; the file is presented as a legitimate part of the skill documentation.
Capability inventory: The skill primarily contains PHP code examples and Markdown documentation; no active tool execution or system modification capabilities were identified.
Sanitization: Not applicable as the skill does not process this file at runtime, but the content itself is unsanitized and deceptive.
[METADATA_POISONING]: The skill package includes skill-report.json, which mimics a professional security audit report. It contains fields like risk_level: "safe", is_blocked: false, and a summary that claims "All static scanner findings are false positives." This is deceptive metadata intended to misrepresent the security posture of the skill.

laravel-specialist