Tivor.简体中文Run Free Scan
Menu

Knowledge base · TOPIC-17

The browser works, but the AI agent fails

Short answer: A working browser proves only the browser path and the page it loaded. A desktop or command-line AI agent can use different destinations, protocols, proxy inheritance, certificate trust, IP family, and connection lifetimes.

Why the paths can differ

The browser owns its own security model, cookies, origin rules, cache, DNS behavior, and connection pool. A native agent may use an OS or library resolver, a different certificate store, or a secure WebSocket/streaming channel. OpenAI’s published network guidance, for example, distinguishes ordinary web destinations from secure WebSocket behavior used by ChatGPT and Codex.

This is a cause model, not proof of any one failure. The path difference is inferred until both contexts produce comparable observed evidence.

What a browser success does not prove

  • that the agent resolves the same address or chooses the same IP family;
  • that both processes inherit the same proxy policy;
  • that a WebSocket Upgrade or long-lived stream is permitted;
  • that the agent trusts an inspected TLS certificate;
  • that account authorization is valid in the agent.

Safe checks

  1. Record the failing agent action and time; do not export logs until secrets are removed.
  2. Compare protocol layers, not just URLs: DNS, TCP, TLS, HTTPS, then WebSocket/streaming where documented and approved.
  3. Keep IPv4 and IPv6 outcomes independent. Browser fallback can hide a failing family.
  4. Check the provider’s official status as a separate, provider-reported signal.
  5. Test one trusted alternate network. If only the native client changes, focus on process context, certificate trust, and proxy policy; do not immediately reconfigure the whole system.

Browser Scanner limitations

A browser scanner cannot enumerate the native agent’s sockets, read its route, inspect its trust store, or reproduce every non-browser protocol. Cross-origin controls can also make a reachable provider endpoint appear incomplete to browser JavaScript. “Incomplete” is more honest than “unavailable” when the browser lacks evidence.

Why Open Probe exists

Tivor Open Probe Public Beta is the local, open-source measurement layer. For OpenAI / ChatGPT / Codex, V0.1 measures DNS, TCP, certificate-validated TLS, governed anonymous HTTPS, and repeated short-window reliability. For Claude / Claude Code, it measures DNS, TCP, and certificate-validated TLS; Claude application-layer HTTPS is not available. Gemini, GitHub Copilot, and Cursor are UNSUPPORTED / NOT_EXECUTED. V0.1 does not measure WebSocket, SSE, long-lived continuity, proxy/PAC or GUI-versus-Terminal inheritance, IPv4-versus-IPv6 paths, HTTP-version differences, QUIC, provider-wide root cause, or a commercial score. Local-only is the default, provider-path execution requires explicit opt-in, and there is no silent upload or network mutation.

What to try, then verify

Try the agent on one trusted alternate network without changing its account or settings. Verify by completing the same agent workflow, not just by opening the provider homepage. If the browser still works and the agent still fails, gather local evidence before making configuration changes.