skills/shkarupa-alex/benz/fuel-watch/Gen Agent Trust Hub

fuel-watch

Pass

Audited by Gen Agent Trust Hub on Oct 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPERSISTENCEDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill extracts text data (titles, addresses, statuses) from external websites like Yandex Maps and 2GIS. Processing untrusted third-party content for reports creates a potential ingestion surface for malicious instructions.
  • Ingestion points: yandex.mjs, gdebenz.mjs, twogis.mjs, and benzonavt.mjs perform data extraction from remote websites via a browser tool.
  • Boundary markers: Present; the SKILL.md instructions explicitly tell the agent to treat the generated report as an immutable block and not to rephrase or interpret the content, creating a strong logical boundary.
  • Capability inventory: The skill possesses capabilities for local file system access (for history and state), CLI execution (agent-browser), and process management (ps).
  • Sanitization: String normalization is applied in normalize.mjs, and extracted addresses are escaped for Markdown in render-report.mjs to prevent structural breakage.
  • [PERSISTENCE]: The skill includes a deployment script (deploy/linux/install-systemd-user.sh) that sets up a systemd user timer and service for periodic background data collection.
  • This mechanism is a primary, documented feature of the skill to maintain a history of fuel availability without requiring an active user or agent session.
  • It is user-initiated and requires running a specific setup script in a Linux environment.
  • [DYNAMIC_EXECUTION]: The skill generates and executes JavaScript strings within the browser environment (via evalJson) to perform data extraction from targeted websites.
  • These scripts are defined within the skill's source code and are used to interact with the DOM of external sites.
  • The logic includes integrity checks, such as hashing the adapter source code during build time to ensure code continuity.
  • [COMMAND_EXECUTION]: The skill executes external command-line tools as part of its operational cycle.
  • It invokes agent-browser to perform headless or headed web scraping operations.
  • It invokes ps and kill (via browser-runner.mjs) to manage its browser daemons and cleanup orphaned processes, using strict PID and command-line signature verification to ensure it only signals its own processes.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 15, 2026, 07:17 PM