The Agent Skills Directory

[PROMPT_INJECTION]: No malicious prompt injection patterns were found. The use of instructional terms like 'IMPORTANT' and 'CRITICAL' is strictly contextual for medical research methodology and quality control.
[DATA_EXFILTRATION]: No data exfiltration patterns detected. The script cohort_overlap_check.py uses urllib.request to communicate with the official NCBI PubMed E-Search API (eutils.ncbi.nlm.nih.gov), which is a well-known scientific service. No sensitive local files (e.g., credentials or SSH keys) are accessed.
[REMOTE_CODE_EXECUTION]: No remote code execution patterns or unsafe downloads-then-execute chains found. The skill executes locally provided Python and R scripts for statistical analysis.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted research data (PDFs converted to Markdown), which is an inherent risk category. However, the workflow is architected with multiple 'fail-closed' gates, mandatory human-in-the-loop verification steps (as seen in extract_assist.py), and deterministic reconciliation scripts that greatly reduce the risk of an agent acting on instructions embedded in research papers.
[OBFUSCATION]: No obfuscated code, Base64-encoded strings, zero-width characters, or homoglyph attacks were detected in the scripts or reference metadata.
[DYNAMIC_EXECUTION]: The skill scripts dynamically import yaml to process configuration files, but do not use unsafe deserialization (e.g., pickle) or eval/exec on untrusted inputs.
[PRIVILEGE_ESCALATION]: No commands related to privilege escalation (e.g., sudo) or unauthorized file system permissions modifications were found.

meta-analysis