ai-data-remediation
Warn
Audited by Gen Agent Trust Hub on Sep 13, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill generates Python lambda functions at runtime using a language model as described in Step 3 and executes them across data clusters in Step 4. While the skill specifies using 'AST Safety Verification' to block dangerous calls like 'import', 'exec', or 'eval', runtime code generation from dynamic prompts remains a notable attack surface.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests potentially untrusted 'anomalous records' to identify patterns and generate fix logic in Step 2 and 3. If an attacker can influence these records, they might manipulate the logic synthesis process.\n
- Ingestion points: Failed rows from external data sources are ingested into an isolation buffer (SKILL.md, Step 1).\n
- Boundary markers: The skill mentions prompt-constrained output but does not specify explicit boundary delimiters for untrusted data within its instructions.\n
- Capability inventory: The system executes generated code within a Python environment after AST parsing (SKILL.md, Step 4).\n
- Sanitization: The skill employs AST static analysis to reject unauthorized node types and dangerous code constructs like imports or OS calls.
Audit Metadata