skills/mukul975/anthropic-cybersecurity-skills/assessing-vector-and-embedding-weaknesses/Gen Agent Trust Hub
assessing-vector-and-embedding-weaknesses
Pass
Audited by Gen Agent Trust Hub on Aug 3, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill provides legitimate security auditing capabilities aligned with OWASP LLM08 (Vector and Embedding Weaknesses) and MITRE ATLAS framework techniques. The provided scripts perform mathematical similarity checks and database queries to identify potential data leakage or poisoning risks in AI systems.\n- [DATA_EXPOSURE]: The skill interacts with vector database endpoints (e.g., Qdrant, Pinecone) as specified by the user. The
scripts/agent.pyscript retrieves results and prints similarity scores to the console for analysis. There is no evidence of data exfiltration to unauthorized remote servers.\n- [COMMAND_EXECUTION]: The skill utilizes standard Python libraries for AI and data science tasks. No suspicious use ofos.system,subprocess, or dynamic code execution wrappers was found.\n- [PROMPT_INJECTION]: The skill includes a defensive scanner (cmd_injectioninagent.py) designed to identify and flag indirect prompt injection markers within retrieved database chunks. This is an auditing feature intended to protect systems rather than exploit them.
Audit Metadata