Bug Description
I'm using the Fable model for daily security hardening on our SaaS project, and I keep getting downgraded to Opus by the safeguard regardless of what I do. This is my own application. All I'm asking for is an audit of my security measures — to find whether anyone can bypass them so I can fix the gaps. The model has authenticated access to my server, codebase, and database, which confirms this is my project and that we're doing legitimate DevOps work. Despite that, it refuses, which is disruptive.
I understand that more capable models carry more risk, but the safeguard needs to be context-aware. I assume the mechanism is a middleware that reroutes flagged prompts to a lower-tier model instead of letting them reach the stronger one. That's a sound design. My request is that the middleware evaluate context before acting — proof of ownership, authenticated infrastructure access, stated defensive intent — rather than downgrading prompts blindly.
Specifically, it should:
- Distinguish authorized security work (owner auditing their own system) from actual misuse.
- Weight signals of legitimacy: authenticated server/database/codebase access, project ownership, consistent defensive framing.
- Downgrade or block only when context genuinely indicates risk, not on keyword matches like "bypass" or "audit."
The current behavior blocks legitimate defensive work and produces a worse experience than a smarter gate would.
Environment Info
- Platform: darwin
- Terminal: Apple_Terminal
- Version: 2.1.198
- Feedback ID: 68bc10fa-8298-47c4-a3ab-df97211265d7
Errors
Bug Description
I'm using the Fable model for daily security hardening on our SaaS project, and I keep getting downgraded to Opus by the safeguard regardless of what I do. This is my own application. All I'm asking for is an audit of my security measures — to find whether anyone can bypass them so I can fix the gaps. The model has authenticated access to my server, codebase, and database, which confirms this is my project and that we're doing legitimate DevOps work. Despite that, it refuses, which is disruptive.
I understand that more capable models carry more risk, but the safeguard needs to be context-aware. I assume the mechanism is a middleware that reroutes flagged prompts to a lower-tier model instead of letting them reach the stronger one. That's a sound design. My request is that the middleware evaluate context before acting — proof of ownership, authenticated infrastructure access, stated defensive intent — rather than downgrading prompts blindly.
Specifically, it should:
The current behavior blocks legitimate defensive work and produces a worse experience than a smarter gate would.
Environment Info
Errors