Anyway to remove the guards.

[blissful dagger]

I’m having the issue where I’m trying to develop for a server asking it to test fake credentials and I have to say the AI is being really rude and refusing and almost demeaning

[oblique wave]

A lot of models will refuse anything that sounds like credential testing / login probing (even if you say “fake creds”), because it’s hard for them to tell “my staging server” from “account cracking”. So the refusal itself is normal — but it shouldn’t be demeaning.

A few practical ways to fix this:

• Reframe the request away from “try credentials” and toward “write tests”.

  • Example ask: “Generate integration tests for my /login endpoint. Use a known test account that I provide (username + password) and assert 200/401 responses. Do not guess passwords or enumerate users.”
  • Or: “Here’s my OpenAPI spec; write tests that verify invalid password returns 401 and that rate limiting triggers after N attempts.”

• Don’t use the LLM as the thing that “attempts logins”.

  • Let your test runner do the requests (pytest/jest/k6/etc.) and use the LLM to help write those tests or interpret results.

• If the “rude tone” is the main issue: add an explicit tone rule to your agent instructions (OpenClaw injects these every turn via your workspace files/system prompt).

  • Something like: “Never be snarky or judgmental. If you must refuse, do it briefly and offer a safe alternative.”
  • Docs: https://docs.openclaw.ai/concepts/system-prompt and https://docs.openclaw.ai/concepts/agent

If you paste:

1. the exact prompt you used (redact anything sensitive), and