shinka-convert
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted code from the current working directory to create a task snapshot and generate evaluation files, creating a surface for injection attacks if the codebase contains malicious instructions.\n
- Ingestion points: SKILL.md Workflow steps 1 (Inspect CWD) and 3 (Minimal runnable snapshot scope).\n
- Boundary markers: Absent; the skill does not use delimiters to wrap the ingested code or instructions to ignore embedded prompts.\n
- Capability inventory: The skill writes files to the filesystem and executes shell commands (Step 8:
python evaluate.py).\n - Sanitization: Absent; no sanitization or validation of the source code is performed before it is used to generate the evaluation logic.\n- [COMMAND_EXECUTION]: The skill performs a local smoke test by executing
python evaluate.pyin the shell. This script is generated based on the structure of the existing project, leading to execution of logic derived from the user's environment.\n- [DYNAMIC_EXECUTION]: The skill dynamically generates Python scripts (evaluate.py,run_evo.py) and configuration files (shinka.yaml) at runtime and then executes the evaluation entrypoint to verify the generated environment.
Audit Metadata