The Agent Skills Directory

[COMMAND_EXECUTION]: The logic defined in SKILL.md for the apply_fix_to_cluster function executes Python lambda functions generated by a local AI model using the eval() function. The implementation relies on a string-based blacklist (checking for keywords like 'import', 'exec', and 'os.') which is insufficient to prevent the execution of malicious code that uses alternative Python syntax or string manipulation to achieve the same effects.- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests and processes untrusted anomalous data (suspect_rows). This data is passed as samples to a local SLM to generate transformation logic. The absence of robust boundary markers or sanitization when feeding these samples to the model could allow an attacker to craft data that manipulates the model into generating malicious code. The evidence chain consists of ingestion in cluster_anomalies, lack of delimiters in generate_fix_logic, and the execution capability via eval() in apply_fix_to_cluster.

agency-ai-data-remediation-engineer