The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses persona-based role-play instructions to guide the agent's behavior. It includes a 'Stop (only once)' directive to ensure a disclaimer is shown only upon the first activation, and defines 'EXIT TRIGGER' phrases to return to normal mode. These are standard persona management techniques and do not attempt to bypass safety guidelines.
[REMOTE_CODE_EXECUTION]: The README provides installation instructions for the user to run npx skills add or git clone. These are standard procedures for the Agent Skills ecosystem and target known repositories from the author and vercel-labs. The agent itself is not instructed to execute untrusted remote scripts.
[DATA_EXFILTRATION]: No evidence of unauthorized data access or external transmission of sensitive information was found. The skill mandates the use of search tools to verify technical facts, which is a standard operational procedure for RAG-based agents.
[INDIRECT_PROMPT_INJECTION]: The skill processes external information via a 'WebSearch' protocol to keep its technical advice current. While this creates a surface for potential indirect injection from untrusted web content, it is a standard risk for all search-enabled AI agents and the skill includes specific role-play boundaries to mitigate unintended behavior.
[OBFUSCATION]: A thorough scan of the research files and skill instructions revealed no base64, hex, zero-width characters, or homoglyph-based obfuscation. All documentation and links are in plain text.

ilya-sutskever-perspective