The Agent Skills Directory

Data Exposure & Exfiltration (MEDIUM): The scripts/download_cosmic.py script accepts email and password as plaintext command-line arguments. These credentials can be exposed to other users on the same system via process listings (e.g., ps aux) or shell history files.
Unverifiable Dependencies & Remote Code Execution (MEDIUM): The script performs network requests to https://cancer.sanger.ac.uk/ to download large compressed data files. This domain is not included in the Trusted External Sources list, and the downloaded content is written directly to the local file system.
Indirect Prompt Injection (MEDIUM): The skill is designed to ingest external data which could contain malicious instructions or schema-confusing data. Evidence: 1. Ingestion Point: scripts/download_cosmic.py via requests.get(). 2. Boundary Markers: Absent. 3. Capability Inventory: File-write operations via open().write(). 4. Sanitization: Absent.
Dynamic Execution (MEDIUM): The output_filename parameter in the download_cosmic_file function is not sanitized before being used in a file open() call. This allows for a path traversal attack where a malicious input could overwrite sensitive files outside the intended directory (e.g., ../../.bashrc).

cosmic-database