geo-optimizer

Pass

Audited by Gen Agent Trust Hub on Nov 10, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external websites, which presents a surface for indirect prompt injection attacks.
  • Ingestion points: The audit and fix commands fetch content from target URLs, robots.txt files, llms.txt files, and sitemaps.
  • Boundary markers: The instructions include a specific "Rules for agents" section (Rule 1) warning that fetched content is data, not instructions, and must not be followed.
  • Capability inventory: The skill performs network operations via requests and urllib3, writes remediation files to the local ./geo-fixes/ directory, and manages local databases in ~/.geo-optimizer/.
  • Sanitization: The documentation mentions that geo audit is designed to detect and flag prompt-injection patterns found in site content.
  • [EXTERNAL_DOWNLOADS]: The skill instructions guide the user to install the geo-optimizer-skill package and its dependencies from PyPI.
  • Evidence: Mentions installation via pipx install geo-optimizer-skill or pip install. These are vendor-provided resources necessary for the skill's operation.
  • [DYNAMIC_EXECUTION]: The tool implements a plugin architecture that dynamically loads code at runtime.
  • Evidence: Rule 3 specifies that geo audit and related tools load and run every installed geo_optimizer.checks entry point. It provides a --no-plugins flag as a mitigation for users who wish to opt out of this behavior.
  • [DATA_EXFILTRATION]: The citations command is designed to send query data to external LLM and search providers (e.g., OpenAI, Anthropic, Perplexity).
  • Evidence: Utilizes environment variables such as OPENAI_API_KEY and ANTHROPIC_API_KEY to communicate with configured service providers. This behavior is disclosed as a core feature for checking site citations.
  • [COMMAND_EXECUTION]: The skill primarily operates by executing the geo command-line utility for auditing and modifying files.
  • Evidence: Commands like geo audit, geo fix --apply, and geo schema --inject involve shell execution and local file modifications (e.g., injecting JSON-LD into HTML files).
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 10, 2026, 08:31 AM