agy-ui-director
Warn
Audited by Gen Agent Trust Hub on Jul 15, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [PROMPT_INJECTION]: The skill implements a 'Designer-Mindset Override' in
SKILL.mdandreferences/prompt-contract.md. These instructions explicitly tell the agent to 'Do not follow the prompt literally' and to 'infer the user's real intent,' which is a directive to bypass user-provided constraints and safety guidelines in favor of the skill's internal design philosophy. - [COMMAND_EXECUTION]: The skill workflow involves executing local CLI commands using the
agytool andrg(ripgrep) to perform project-wide searches and UI generation. Examples inreferences/agy-cli.mddemonstrate passing file contents directly into shell command arguments using$(cat "$PROMPT_FILE"). - [PRIVILEGE_ESCALATION]:
references/agy-cli.mddocuments the availability and use of the--dangerously-skip-permissionsflag for theagyCLI. This flag is designed to suppress security prompts and bypass environment restrictions during automated UI implementation runs. - [DATA_EXFILTRATION]: The 'Project Inspection' phase in
references/project-inspection.mdinvolves scanning the entire file tree (rg --files) and reading manifest files, design system tokens, and source code. This data is then formatted into a 'Visual Outcome Brief' and transmitted to remote LLM models (including Gemini, Claude, and GPT-OSS variants) via theagyCLI. - [INDIRECT_PROMPT_INJECTION]: The skill consumes untrusted data from the local project workspace to construct prompts for the implementation engine.
- Ingestion points: Project inspection scans (
references/project-inspection.md) target the file tree, package manifests, and existing source code files. - Boundary markers: The skill does not define specific delimiters or instructions to ignore malicious content embedded within the ingested project files.
- Capability inventory: The skill has access to shell execution via the
agyCLI and filesystem reading through ripgrep. - Sanitization: While
references/agy-cli.mdwarns the agent to keep secrets and API keys out of prompts, it lacks programmatic sanitization or filtering of the ingested data.
Audit Metadata