scientific-critical-thinking
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute a shell command (python scripts/generate_schematic.py) using natural language input directly from the user. This creates a risk of command injection if the user includes shell metacharacters that are not properly sanitized before the command is run using the Bash tool.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze external scientific documents, which are untrusted sources. It lacks instructions for using boundary markers or sanitizing content, and given the skill's access to Bash and file writing tools, this presents a vulnerability to indirect prompt injection.
- Ingestion points: Evaluation of research methodology, scientific claims, and research papers (SKILL.md).
- Boundary markers: None identified.
- Capability inventory: Bash, Read, Write, and Edit tools (SKILL.md).
- Sanitization: None identified.
- [METADATA_POISONING]: The skill contains instructions to suggest and redirect users to an external website (www.k-dense.ai). There is a discrepancy between the system-reported author (wu-yc) and the metadata author (K-Dense Inc.), and the domain is not listed as a known resource for the author, potentially leading to misleading service promotion.
Audit Metadata