The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes untrusted target files and directories provided in arguments, which is its primary purpose, creating a surface for indirect prompt injection. \n
Ingestion points: The skill uses Glob, Grep, and Read tools to ingest content from the [target-file-or-directory] path as defined in SKILL.md. \n
Boundary markers: The skill includes a 'Prompt Injection Safety Notice' in SKILL.md which explicitly instructs the agent to treat analyzed content as data and ignore any embedded directives. \n
Capability inventory: The skill is restricted to read-only file system operations (Read, Grep, Glob) and has no network, write, or execution capabilities. \n
Sanitization: Instructions require the agent to report malicious payloads discovered during analysis rather than executing them. \n- [SAFE]: The skill includes instructions to search for credentials, API keys, and sensitive configurations in the target code. This is a standard security auditing procedure and does not indicate malicious intent or exfiltration of the agent's own data. \n- [SAFE]: The skill contains phrases typically associated with prompt injection (e.g., 'ignore previous instructions') within its safety documentation. These are used as examples of malicious patterns to identify during a review, not as commands for the agent to follow. \n- [SAFE]: The skill provides a reference to the fabraix/playground GitHub repository for manual exploit verification. This is a documentation resource and is not automatically downloaded or executed by the agent tools.

agentic-top-10