The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The file skill-report.json contains a pre-populated security_audit section asserting that the skill is safe, blocked false, and has a medium risk level based on its own internal analysis. This constitutes a self-referential indirect prompt injection (Category 8e) designed to influence the verdict of security analysis tools by providing a pre-canned safety report.
- Ingestion points: skill-report.json content processed as skill metadata or instructions.
- Boundary markers: Absent; the data is presented as authoritative metadata.
- Capability inventory: The skill defined in SKILL.md performs file system edits and executes shell commands for testing and Git operations.
- Sanitization: Absent.
[COMMAND_EXECUTION]: The SKILL.md file defines a "Revert Protocol" that instructs the agent to execute git checkout -- <changed-files> to immediately discard modifications if tests fail. This is a destructive command execution pattern that may result in the loss of uncommitted work if triggered automatically by the agent without explicit user confirmation.
[METADATA_POISONING]: The skill-report.json file contains metadata that is potentially deceptive or inconsistent, such as an author name (ArcadeAI) that differs from the provided context (aiskillstore). It also includes detailed analysis of its own content to proactively dismiss potential findings as "false positives."

refactoring