The Agent Skills Directory

[DYNAMIC_EXECUTION]: The reference for AI data remediation defines a workflow where a local AI model (Ollama) generates Python lambda functions that are subsequently executed using the eval() function.
File: references/engineering-ai-data-remediation-engineer.md
Evidence: The implementation of apply_fix_to_cluster uses transform_fn = eval(fix['transformation']) to apply AI-generated logic to data clusters.
Risk: Dynamically executing code generated by an LLM is a high-risk pattern, particularly when the LLM's output is derived from untrusted input data. Although a blacklist-based safety check is described, such filters are often insufficient to prevent sophisticated injection attacks.
[INDIRECT_PROMPT_INJECTION]: The system is designed to process 'anomalous' data through an AI model, which exposes a surface for indirect prompt injection attacks.
Ingestion Points: The suspect_rows parameter in references/engineering-ai-data-remediation-engineer.md is sourced from data pipelines and is potentially untrusted.
Capability Inventory: The skill possesses the capability to execute code via eval(), which can be leveraged if an injection succeeds.
Sanitization: A basic blacklist (blocking import, exec, eval, os., subprocess) is implemented as a safety gate, but it does not account for all methods of achieving malicious execution in Python.
[COMMAND_EXECUTION]: The DevOps and Incident Response references include numerous examples of powerful shell commands for managing cloud infrastructure and containers.
Files: references/engineering-devops-automator.md, references/engineering-incident-response-commander.md.
Evidence: Frequent use of kubectl, terraform, aws, and docker commands within the instructional content and code snippets.

engineering