kdp-cover-wrap
Fail
Audited by Gen Agent Trust Hub on Sep 30, 2026
Risk Level: HIGHDATA_EXFILTRATIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [DATA_EXFILTRATION]: The skill exposes absolute file paths belonging to the author's local environment, disclosing the system's username and private directory structure. Accessing files in a user's home directory is a significant boundary violation.
SKILL.mdinstructs the agent to access a private rule file at/Users/charliedeist/Root/.claude/rules/writing-quality.md.scripts/html_wrap_to_pdf.mjshardcodes a Node.js module path:/Users/charliedeist/.npm-global/lib/node_modules/agent-browser/node_modules/playwright-core/index.mjs.scripts/html_wrap_to_pdf.mjshardcodes the browser executable path:/Applications/Google Chrome.app/Contents/MacOS/Google Chrome.- [COMMAND_EXECUTION]: The orchestration script
scripts/pipeline.pyutilizessubprocess.runto execute multiple internal Python scripts. While it uses list-based arguments to prevent simple shell injection, it establishes a capability for the skill to spawn and control new processes on the host system. - [DYNAMIC_EXECUTION]: The skill performs runtime manipulation of binary files and executes external software through computed or hardcoded absolute paths.
scripts/html_wrap_to_pdf.mjsperforms direct byte-level patching of generated PDF files to modify/MediaBoxdimensions using string replacement on the file buffer.scripts/html_wrap_to_pdf.mjslaunches and controls a headless Google Chrome instance via Playwright to render artboards.scripts/html_wrap_to_pdf.mjsimports JavaScript modules from an absolute local filesystem path outside the skill's distribution folder.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied data which is interpolated into prompts for generative models and passed to composition scripts, creating a surface for injection attacks.
- Ingestion points: User-provided book titles, author names, creative direction, and back cover copy are processed via
scripts/pipeline.py,scripts/gen_openai.py, and HTML templates. - Capability inventory: Spawns sub-processes (
subprocess.run), performs file system writes (Pillow), executes browser sessions (Playwright), and interfaces with external generative AI APIs (OpenAI). - Boundary markers: The skill suggests a "Safe-zone preamble" in
prompts/leaf-cover.mdto guide the image model's layout behavior, but lacks formal delimiters or warnings to prevent instruction override from the user-provided text content. - Sanitization: Inputs such as the back cover blurb are inserted into HTML templates and passed as CLI arguments without explicit sanitization or escaping mechanisms.
- [EXTERNAL_DOWNLOADS]: The skill instructions describe procedures for fetching external content from well-known services.
SKILL.mdfetches handoff bundles and design artifacts from Anthropic's design API (api.anthropic.com).
Recommendations
- AI detected serious security threats
Audit Metadata