magic-data-synthesis
Warn
Audited by Gen Agent Trust Hub on Jul 3, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The script
scripts/batch_synthesize.pyemploys the Pythoneval()function to execute logic defined in theexpressionstrategy. Although it implements a basic sandbox by clearing__builtins__, this technique is frequently bypassable and represents a risk if the configuration containing the expression is sourced from an untrusted entity. - [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection.
- Ingestion points: Data is ingested from CSV, Parquet, JSON, and JSONL files via
scripts/batch_synthesize.pyandscripts/generate_column.py. - Boundary markers: The system uses standard string interpolation in
synthesis_prompt_builder.pywithout robust separators to distinguish data from instructions. - Capability inventory: The skill can perform arbitrary Python execution (via
eval()inbatch_synthesize.py), shell command execution through thedata-designerCLI, and network operations for LLM API calls. - Sanitization: Basic sanitization is performed in
synthesis_utils.py(stripping newlines and truncating values), but it does not prevent the execution of malicious instructions embedded within the processed text.
Audit Metadata