The Agent Skills Directory

[DATA_EXFILTRATION]: The 'Jev step judge' feature (controlled by ORK_EXPECT_JEV) transmits application context to a third-party API endpoint (https://api.typesafe.ai/v1/systemone). Data sent includes current step goals, verification results, and interactive element metadata (ARIA roles and names). While the skill implements a redaction mechanism in scripts/jev_shadow.py to filter common secret patterns (tokens, keys, passwords), structural application context is still transmitted externally.\n- [REMOTE_CODE_EXECUTION]: The session recording functionality described in references/rrweb-recording.md involves injecting a remote script from cdn.jsdelivr.net and executing it within the browser context using eval. Although fetching from a well-known CDN, this represents a remote code execution vector in the browser environment.\n- [EXTERNAL_DOWNLOADS]: The skill documentation recommends installing external tools such as agent-browser via npm and fetches library code from public CDNs during runtime execution.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from browser ARIA trees and git diffs to generate automated test plans via AI agents. This creates an attack surface where malicious page content could influence the agent's test plan generation.\n- [COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool for system operations, including git diff analysis, file fingerprinting with SHA-256, and executing local Python scripts for data processing.\n- [DYNAMIC_EXECUTION]: The skill generates and executes logic at runtime, including browser-side scripts for session capture and Python scripts used for judge integration and UI analysis.

expect