[Bug] Safety classifier blocks authorized security-testing content in assistant output, but not when loaded via Skill tool

Status Open
Reported on v2.1.288
Maintainer reply None cached
Activity 2 comments · opened Oct 3, 2026

Bug Description
Using Opus 5.5 on a Max plan, I get safeguard blocked when using even pretty harmless QA prompts. For example, from a completely new session:

Let's change gears and do some testing instead of planning. Our pre-flight list is getting pretty long. 
We'll start with (Linear ticket ID redacted). It's familiar ground; the second to last comment in the 
ticket was from me.

I even get blocked when I try to ask Opus for help with this issue: still on Opus 5.5, I sent it this prompt:

Need your help with something. I issue this innocuous, very QA prompt:
  
  Let's change gears and do some testing instead of planning. Our pre-flight list is getting pretty long. We'll start with AE-79. It's familiar ground; the second to last comment in the ticket was from me.
  
  I get:
  
   [Pasted text #3 +7 lines]  <-- that's the safeguard message
  WTF??

Switching to Opus 5 got me past this.

Environment Info

  • Platform: darwin
  • Terminal: Apple_Terminal
  • Version: 2.1.288
  • Feedback ID: 3211b997-4a8e-474b-b684-f004f808b137

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗