Security audit
Security checks for vulnerabilities and agentic risk
This skill is a local RAG helper toolkit whose file access and writes are user-directed and purpose-aligned, with some reliability and data-handling caveats.
Install only if you are comfortable with a local tool reading your chosen document folders and writing derived chunks, results, and index files. Use dedicated working directories, do not point outputs at important folders, and treat any future use of real hosted embedding APIs as a separate privacy decision because document chunks or queries could be sent to those providers.
The skill recommends paid embedding providers without warning that document chunks or query text may be sent to third-party APIs for processing. Because RAG inputs often contain proprietary or regulated data, this omission can lead users to unknowingly exfiltrate sensitive dataset contents outside their trust boundary.
The skill instructs users to write chunk outputs to arbitrary user-specified directories but does not warn that running these commands can create or overwrite files in those paths. In an agent or automation context, this increases the chance of unintended filesystem modification, data clobbering, or writing sensitive derived data into insecure locations.
The module and function documentation present this as an embedding benchmark, but the implementation uses a mock similarity routine based on token overlap rather than generating embeddings or invoking the listed models. This is an active contradiction about what the code actually evaluates, which can mislead users about the skill's real behavior.
The --delete operation performs a persistent file-backed deletion through store.delete_document() without any confirmation prompt, warning comment, or prior disclosure beyond the terse CLI help text. Unlike --clear, which asks for confirmation, this destructive action can remove indexed data immediately and silently except for the post-action print.
This code writes benchmark results to a user-specified path, which modifies the local filesystem. Although it prints after saving, there is no prior warning, confirmation, or explicit disclosure that an existing file may be overwritten when --output is used.
This is a manifest file, so vague-trigger review applies. The description explains the skill's domain and capabilities but does not define when it should be invoked, what exact requests should trigger it, or any exclusions, which can lead to overly broad matching for general RAG-related requests.
No suspicious patterns detected.
