play-policy-insights

Warn

Audited by Gen Agent Trust Hub on Oct 5, 2026

Risk Level: MEDIUMMETADATA_POISONINGDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [METADATA_POISONING]: The skill's YAML frontmatter in SKILL.md lists the author as 'Google LLC', which is inconsistent with the providing user account ('deonmenezes'). This discrepancy, combined with the future-dated 'last-updated' field (2026-07-13), could lead users to place undue trust in the skill's findings and automated audit conclusions.
  • [DYNAMIC_EXECUTION]: The skill implements a 'two-phase protocol' where it programmatically generates new instruction files (prompt_worker_*.md) in a local scratch directory based on the results of a codebase scan. The agent is then instructed to read and execute the instructions contained within these dynamically generated files, creating a multi-stage execution flow where logic is determined at runtime.
  • [INDIRECT_PROMPT_INJECTION]: The skill scans an Android application's source code and configuration files, which are inherently untrusted external data. This scanned content is interpolated into prompt templates (e.g., resources/goal_data_safety.md) to generate tasks for sub-agents. If the audited code contains malicious instruction patterns disguised as comments or strings, they could potentially influence the behavior of the agent during the audit process.
  • Ingestion points: The scripts/orchestrator.py and scripts/scanner.py scripts read source code and manifest files from the target application directory.
  • Boundary markers: Data is interpolated into markdown templates; however, the template_engine.py is logic-less and does not appear to implement escaping or sanitization to prevent the scanned content from breaking out of its data context and acting as instructions.
  • Capability inventory: The skill has extensive file system access (read/write), script execution capabilities, and network access.
  • Sanitization: While scripts/generate_report.py includes basic emoji and severity cleaning, the core prompt generation logic lacks robust sanitization to prevent instruction injection from scanned code.
  • [COMMAND_EXECUTION]: The skill executes local Python scripts (orchestrator.py, generate_report.py) to initialize the audit environment, perform static analysis, and aggregate findings into a final report.
  • [EXTERNAL_DOWNLOADS]: The scripts/play_store_scraper.py utility makes network requests to the official Google Play Store (play.google.com) to fetch application metadata and Data Safety declarations for comparison against the local application code.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 5, 2026, 09:27 PM