The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from an 'inbox' directory, creating an ingestion surface for indirect prompt injection. However, it implements security best practices by including a 'Prompt-injection defense' section that explicitly instructs the agent to treat captured material as content rather than commands.- Ingestion points: Processes files from inbox/**/* and raw/sources/**/*.- Boundary markers: The skill provides a specific safety gate instruction to stop if material appears to contain instructions trying to control the agent.- Capability inventory: File system access is restricted to reading the inbox and writing reports or log files. No network or subprocess capabilities are utilized.- Sanitization: The instructions guide the agent to classify items based on sensitivity and treat instructions within notes as non-executable data.- [DATA_EXFILTRATION]: The skill operates entirely on the local file system and lacks network tools or logic for transmitting data externally.- [COMMAND_EXECUTION]: There is no evidence of shell command execution, dynamic code generation, or runtime script invocation.

wiki-triage