hyper-browsing

Pass

Audited by Gen Agent Trust Hub on Nov 19, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to fetch untrusted content from the web, which is a known attack surface. However, it implements proactive defenses in rules/access-policy.md (§6) and rules/evidence-and-report.md (§5) that explicitly instruct the agent to treat any text resembling commands or instruction overrides (e.g., "ignore previous instructions") as data to be reported, rather than instructions to be executed.
  • [COMMAND_EXECUTION]: The skill relies on local shell commands such as curl, python3, yt-dlp, and jq to perform its browsing and data extraction tasks. These commands are used according to well-defined recipes in the references/ directory.
  • [EXTERNAL_DOWNLOADS]: The skill may perform automated installations of Python packages like curl_cffi and yt-dlp from PyPI into a task-owned virtual environment when necessary. It also retrieves content from user-specified URLs and reputable third-party services like the Wayback Machine and the Jina Reader API.
  • [DATA_EXFILTRATION]: The skill saves fetched content to the local directory .hyper/browsing/. Its rules/access-policy.md contains strict data-handling policies that prevent the transmission of sensitive user data, such as browser cookies or authorization headers, to external services without explicit consent.
  • [CREDENTIALS_UNSAFE]: The skill includes strong safeguards for user privacy and credentials. It explicitly forbids extracting session data from the user's live browser profile and mandates that users must be present and personally handle any login or CAPTCHA challenges.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 19, 2026, 07:40 PM