The Agent Skills Directory

[COMMAND_EXECUTION]: The skill dynamically executes Python code generated by a local LLM using the eval() function in the apply_fix_to_cluster method.\n
Evidence: transform_fn = eval(fix['transformation']) in SKILL.md.\n
Context: While the skill includes a safety gate that filters for keywords like import, exec, os, and subprocess, the use of eval on model-generated output is a significant security risk, as sophisticated payloads can potentially bypass simple string-based filters via Python introspection.\n- [PROMPT_INJECTION]: The skill processes untrusted data and includes it in LLM prompts, making it susceptible to indirect prompt injection attacks.\n
Ingestion points: The suspect_rows list passed to cluster_anomalies and generate_fix_logic in SKILL.md.\n
Boundary markers: The prompt in generate_fix_logic lacks explicit delimiters for user-provided data samples, increasing the risk of data being interpreted as instructions by the SLM.\n
Capability inventory: The skill possesses the capability to execute code via eval() and perform data operations via pandas.\n
Sanitization: A validation step in generate_fix_logic checks for specific forbidden terms and verifies the lambda prefix.\n- [EXTERNAL_DOWNLOADS]: The skill utilizes the sentence-transformers library, which downloads the all-MiniLM-L6-v2 model from a well-known public repository during initialization.\n
Evidence: model = SentenceTransformer('all-MiniLM-L6-v2') in SKILL.md.

AI Data Remediation Engineer