prompt-scanner
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions direct the agent to construct shell commands by inserting user-supplied text directly into a command string (e.g., agent-sec-cli scan-prompt --text '<prompt_text>'). This design creates a risk of command injection or shell syntax errors if the user input includes single quotes or other shell metacharacters that are not escaped before execution.
- [INDIRECT_PROMPT_INJECTION]: (1) Ingestion points: Untrusted text is ingested through the --text parameter, file inputs, or stdin as documented in SKILL.md. (2) Boundary markers: The skill suggests wrapping text in single quotes as a delimiter, which is a weak boundary that can be subverted. (3) Capability inventory: The skill triggers shell execution of the agent-sec-cli tool. (4) Sanitization: The instructions do not specify a requirement for the agent to sanitize or escape the input before it is interpolated into the shell command.
- [EXTERNAL_DOWNLOADS]: The skill documentation describes fetching machine learning models from modelscope.cn, which is a well-known model registry service.
Audit Metadata