The Agent Skills Directory

[COMMAND_EXECUTION]: In SKILL.md, the 'Target & Disease Literature (OpenTargets)' workflow directly interpolates the ${GENE} variable into a curl shell command string. Unlike other workflows in the skill that use a Python script to URL-encode input, this section lacks sanitization. This creates a vulnerability where malicious input (e.g., using shell command substitution like $(command)) could be executed by the shell environment processing the skill's instructions.
[EXTERNAL_DOWNLOADS]: The skill interacts with several external APIs including ChEMBL (ebi.ac.uk), PubChem (ncbi.nlm.nih.gov), OpenFDA (fda.gov), and OpenTargets (opentargets.org). These are well-known, established scientific and governmental services, and the data retrieved is processed as JSON. Findings related to these trusted sources do not escalate the verdict.
[REMOTE_CODE_EXECUTION]: Multiple instances of curl | python3 were flagged by automated scanners. Detailed technical analysis shows these are false positives for malicious remote code execution; the patterns are used to pipe JSON data from trusted APIs into a local, static Python parser defined within the skill (json.load(sys.stdin)). The code being executed is part of the skill's own source, not code fetched from the internet.

drug-discovery