The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill securely manages the MP_API_KEY required for Materials Project access. It strictly reads the key from environment variables, prohibits its use as a command-line argument, and includes a safe_error_message utility in scripts/_common.py to redact the secret from any exception output. Additionally, scripts/artifact_manifest.py uses a regex to explicitly reject files with names suggesting they contain credentials (e.g., .env, secrets, tokens).
[EXTERNAL_DOWNLOADS]: The skill interacts with the Materials Project API (api.materialsproject.org) via the official mp-api client. This is a well-known scientific service. The skill requires explicit user approval for network operations and includes a dry-run planning mode to disclose network intent before execution.
[COMMAND_EXECUTION]: While the skill utilizes shell and Python tools, it does so through tightly controlled local scripts. These scripts implement significant safety checks, including file size limits (e.g., DEFAULT_MAX_INPUT_BYTES), site count limits for crystal structures, and a non-overwrite policy for all output files using exclusive file opening modes.
[INDIRECT_PROMPT_INJECTION]: The skill protects against malicious content in scientific data files (CIF, JSON) by using a standard-library-first approach for parsing before passing data to scientific libraries. It uses strict JSON loading (rejecting duplicates and non-finite numbers) and schema validation to define trust boundaries.
[PRIVILEGE_ESCALATION]: The skill includes protections against path traversal and symbolic link attacks. The checked_input_file and checked_output_file functions in scripts/_common.py resolve paths strictly and refuse to follow symlinks or allow '..' in output paths.

pymatgen