The Agent Skills Directory

[DATA_EXFILTRATION]: The mandatory 'LIBRARY-FIRST PROTOCOL' instructs the agent to check 'D:\Projects*' to 'EXTRACT and adapt' code. This involves accessing absolute paths on the host's storage that target user project directories, which can lead to unauthorized reading and exposure of sensitive source code, configuration files, or credentials from unrelated projects.
[COMMAND_EXECUTION]: The skill relies on the 'Bash' tool to execute a local script located at 'scripts/multi-model/llm-council.sh'. This script is called with arguments ('query', 'threshold', 'chairman') derived directly from user input, which could facilitate command injection if the underlying shell script does not provide adequate input sanitization.
[EXTERNAL_DOWNLOADS]: The skill references source code and architecture documentation from a well-known public repository and a technology news domain (github.com/karpathy/llm-council and venturebeat.com).
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it ingests untrusted user-supplied queries which are then processed through a multi-model pipeline and a shell execution command.
Ingestion points: User query input via the council command interface.
Boundary markers: No specific delimiters or 'ignore' instructions are defined for the query interpolation.
Capability inventory: Bash execution, file reading, and file writing tools are allowed.
Sanitization: No sanitization or validation of the user query parameter is documented.

llm-council