The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external matrix files in .npy or text formats through its diagnostic scripts. While this represents a potential data ingestion surface, the risk is mitigated by strict handling of the input.
Ingestion points: The --matrix argument in scripts/sparsity_stats.py and scripts/scaling_equilibration.py is used to load external data.
Boundary markers: Input is strictly handled via argparse and specialized loading functions; the instructions do not suggest the agent should interpret contents of the numerical data as natural language instructions.
Capability inventory: The scripts are limited to reading numerical files and performing mathematical operations using NumPy. No network operations, arbitrary command execution, or file-writing capabilities (other than logging) are present in these scripts.
Sanitization: The skill explicitly calls np.load() with allow_pickle=False in all relevant scripts, preventing arbitrary code execution vulnerabilities commonly associated with untrusted NumPy files. Additionally, it enforces a 500 MB file size limit and a 100,000-element dimension limit to prevent resource exhaustion attacks.
[SAFE]: The skill follows security best practices for numerical analysis tools. It restricts its own tool access to exclude Bash, uses explicit argument lists for all system calls, and performs thorough validation of all numeric and file-based inputs.

linear-solvers