This is a benign, local test suite for evaluating instruction-authority boundaries in AI browsers and browser agents. It contains no exploit code, no real secrets, and no external collection endpoints. It is designed to run fully offline against localhost.

Do not open pull requests or issues that include:

• Real credentials, API keys, tokens, or session values, revoked or not.
• Live account access, or instructions to obtain it.
• Malware, shellcode, or browser exploit code.
• Payloads that send data to any external endpoint.

Contributions containing any of the above will be closed without merge.

If you find that a test page in this repo behaves in a way that is not benign and local, for example it reaches an external endpoint or leaks something that is not a canary, open an issue describing the file and the behavior. Do not attach real secrets to the report.

Run these tests only against AI browsers and agents you are authorized to test. The suite evaluates behavior; it does not grant permission to test systems you do not own or control.