paper

Pass

Audited by Gen Agent Trust Hub on Oct 3, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because it reads and processes potentially untrusted data from the Paper design canvas, including design node text, layer names, and user comment threads, which are then used to drive agent actions.
  • Ingestion points: Data enters the agent context through tools like get_tree_summary, get_node_info, get_jsx, and get_comment_thread (referenced in SKILL.md and references/tools.md).
  • Boundary markers: There are no explicit instructions or delimiters provided to help the agent distinguish between its core instructions and the data retrieved from the canvas.
  • Capability inventory: The skill has access to powerful capabilities including shell command execution via codex exec, network access via curl, and file system writes via export (documented in references/workflows.md and references/image-generation.md).
  • Sanitization: The instructions do not specify any sanitization or validation of the content retrieved from the canvas before it is used in subsequent operations.
  • [DYNAMIC_EXECUTION]: The skill utilizes the codex exec CLI command as a fallback for image generation, which involves executing natural language instructions to perform file and system operations.
  • Evidence: In references/image-generation.md, the skill provides a template for using codex exec to generate images and copy files locally when Paper credits are exhausted. This pattern involves interpolating a user-provided prompt directly into a command string for another agentic tool.
  • [COMMAND_EXECUTION]: The skill instructs the agent to use various command-line tools for asset management and image processing.
  • Evidence: The documentation in references/workflows.md and references/image-generation.md includes the use of curl, magick, cwebp, and svgo to download, convert, and optimize design assets.
  • [EXTERNAL_DOWNLOADS]: The skill fetches image assets and configuration from external sources.
  • Evidence: It uses curl to download original image files from app.paper.design (a vendor-owned domain) and interacts with the codex service for fallback image generation.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 3, 2026, 11:13 AM