The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates Python code strings using a local Small Language Model (SLM) and executes them using the eval() function.
Evidence: The function apply_fix_to_cluster in SKILL.md contains the line transform_fn = eval(fix['transformation']).
Risk: While the skill attempts to sanitize the AI-generated string by checking for a subset of dangerous keywords (e.g., import, exec, os), eval() remains a high-risk capability that can be bypassed by sophisticated payloads or library-specific vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data rows and passes samples of these rows into an AI model to generate executable code, creating a significant injection surface.
Ingestion points: Untrusted data enters the agent context via the suspect_rows argument in the cluster_anomalies and generate_fix_logic functions within SKILL.md.
Boundary markers: The prompt construction for the SLM lacks strong delimiters or specialized instruction isolation between the system prompt and the user-supplied data samples.
Capability inventory: The skill possesses the capability to execute dynamic Python code via eval().
Sanitization: The skill relies on basic keyword filtering (import, exec, eval, os., subprocess) to validate the AI's output before execution, which is an incomplete defense against adversarial prompt injection.
[COMMAND_EXECUTION]: The use of eval() facilitates the execution of arbitrary Python expressions, which can be leveraged for broader system interaction depending on the environment.
[EXTERNAL_DOWNLOADS]: The skill mentions the use of local models and libraries such as phi3, Llama-3, and Mistral via Ollama, along with sentence-transformers and ChromaDB for embedding and clustering operations. These are documented as local dependencies for the data remediation workflow.

AI Data Remediation Engineer