The Agent Skills Directory

[COMMAND_EXECUTION]: The audit.py script utilizes subprocess.run to invoke the ripgrep (rg) command-line utility for efficient text searching within repositories. The implementation follows security best practices by using a list of arguments and disabling shell execution, effectively preventing shell injection vulnerabilities.
[REMOTE_CODE_EXECUTION]: The skill incorporates the dspy.RLM framework, which enables the language model to generate and execute Python code dynamically. While this execution is designed to occur within a sandboxed Deno environment and is restricted by prompt instructions, it represents a dynamic code execution pattern where model-derived logic runs on the host system to interact with local tools.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes untrusted source code from the repositories under audit via the read_file and search_pattern tools.
Ingestion points: Untrusted data from the filesystem (at source_root) enters the model's context through file reading and pattern matching tools defined in audit.py.
Boundary markers: The implementation lacks clear delimiters or instructions to the model to ignore potential directives embedded within the audited code content.
Capability inventory: The skill possesses capabilities to write audit reports and metadata to the local filesystem and to execute system binaries like ripgrep.
Sanitization: There is no evidence of content sanitization, filtering, or validation of the source code being analyzed before it is presented to the language model.

security-audit-rlm