The Agent Skills Directory

Data Exposure & Exfiltration (LOW): The script scripts/novelty_check.py makes network requests to api.semanticscholar.org. While this is the intended purpose of the skill, the domain is not on the predefined trusted whitelist.
Indirect Prompt Injection (LOW): The skill ingests untrusted text, such as research areas and idea descriptions, to drive its workflow. Ingestion points: The --idea argument in novelty_check.py and the $0 input in SKILL.md. Boundary markers: Partially present in prompt templates in references/ideation-prompts.md (e.g., triple quotes), but missing for script command-line arguments. Capability inventory: Performs network GET requests to the Semantic Scholar API and can write results to a local file via the --output flag. Sanitization: Search queries are escaped using urllib.parse.urlencode to prevent URL injection.

idea-generation