event-prospecting

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill relies extensively on the Bash tool to orchestrate its pipeline. It executes several local Node.js scripts (recon.mjs, extract_event.mjs, extract_page.mjs, compile_report.mjs) and standard shell commands like mkdir, split, and cat for file management. The compile_report.mjs script specifically uses child_process.execSync to open the final report in the default browser.
  • [DYNAMIC_EXECUTION]: The scripts recon.mjs and extract_event.mjs generate and execute JavaScript strings within the context of the scraped web pages using the browse eval command. This is used to extract structured data from Next.js NEXT_DATA blocks and other DOM elements.
  • [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection because it ingests untrusted content from external event URLs and company homepages.
  • Ingestion points: Step 2 (Recon), Step 3 (Extract), and Step 5/7 (Research) fetch and parse data from arbitrary URLs provided by the user or discovered during research.
  • Boundary markers: The subagent prompts in workflow.md include instructions to ignore cosmetic site features, but do not use explicit delimiters like XML tags or triple-quotes for the ingested HTML/Markdown content.
  • Capability inventory: The skill possesses full Bash access, file write capabilities via heredocs, and network access via the browse CLI.
  • Sanitization: The report generation script (compile_report.mjs) includes robust HTML and JavaScript escaping (escapeHtml, escapeAttr, escapeJsInAttr) to prevent XSS in the generated reports.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 01:47 AM
Security Audit — agent-trust-hub — event-prospecting