page-builder
Fail
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill provides explicit instructions to install an external tool and execute a script from it. In
references/ux-pass.md, it directs the installation of theuipro-clipackage and subsequently instructs the agent to locate and execute a Python script (scripts/search.py) from the resulting installation folder. - [EXTERNAL_DOWNLOADS]: The
references/ux-pass.mdfile contains instructions to download and initialize external software usingnpx uipro-cli init --ai claude. This targets a third-party source not recognized as a trusted vendor or service. - [COMMAND_EXECUTION]: The skill uses shell commands to perform core operations. In
references/ux-pass.md, it specifies runningpython3with a path constructed at runtime to execute an external script, alongside thenpxcommand used for initialization. - [DYNAMIC_EXECUTION]: The skill employs dynamic execution by searching for a file path at runtime based on metadata (
name: ui-ux-pro-max) and executing the found script. Furthermore, Phase 5 inSKILL.mddescribes a dynamic flow where a 'blind judge' subagent is spawned to evaluate content and return structured JSON output. - [INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection.
- Ingestion points: Untrusted data enters the context during Phase 0 (Intake) in
SKILL.md, where user-provided business details, target reader profiles, and search intent are collected for processing. - Boundary markers: The instructions lack any delimiters or markers (e.g., XML tags or clear separators) to isolate untrusted input from the primary system instructions.
- Capability inventory: The skill has shell execution capabilities (
python3,npx) as defined inreferences/ux-pass.mdand file-writing capabilities as implied in the 'CMS split' handoff inreferences/handoff-kit.md. - Sanitization: There is no evidence of input validation, escaping, or filtering of external content before it is interpolated into prompts or used in evaluation passes.
Recommendations
- AI detected serious security threats
Audit Metadata