planf3
Pass
Audited by Gen Agent Trust Hub on Sep 24, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input via the
user_promptvariable to generate implementation plans. These plans include implementation tasks and validation commands ({{VALIDATION_COMMAND}}) which are intended to be executed by the agent in a subsequentBuild Planworkflow. This creates a surface where a malicious prompt could result in the generation and execution of harmful commands.\n - Ingestion points:
user_promptargument inSKILL.md.\n - Boundary markers: The generated HTML plan uses template placeholders but does not include explicit instructions or markers to the agent to ignore or sanitize embedded commands during the build phase.\n
- Capability inventory: The
allowed-toolsincludesBash(uv run:*)andBash(open -a:*). TheBuild Planworkflow (inworkflows/build-plan.md) directs the agent to execute implementation tasks and validation commands defined within the generated HTML plan.\n - Sanitization: No sanitization or validation of the generated commands is performed before execution in the
Build Planworkflow.\n- [DYNAMIC_EXECUTION]: The skill follows a pattern of generating an implementation plan in HTML format (described inworkflows/create-plan.md) and subsequently executing shell commands defined within that plan (described inworkflows/build-plan.md). This constitutes the dynamic generation and execution of code from templates based on user input.\n- [EXTERNAL_DOWNLOADS]: The Python scriptsscripts/generate_gpt_image.pyandscripts/edit_gpt_image.pydeclare dependencies onopenaiandpython-dotenvin their script headers. These packages are downloaded at runtime from the official Python Package Index (PyPI) when executed via theuvtool. These are well-known and standard libraries used for OpenAI API interaction.\n- [COMMAND_EXECUTION]: The skill invokes theBashtool to run local scripts viauv runand to open files in a web browser viaopen -a. Additionally, theBuild Planworkflow explicitly instructs the agent to execute shell commands specified within implementation phases and validation sections of the generated plans.
Audit Metadata