Overly restrictive safety checks cause refusals for benign and routine coding tasks

[syxs22]

Bug Description

OpenClaude is refusing to perform many simple and harmless requests by claiming that its safety directives prevent it from doing so.

The issue is not related to malware, security research, harmful code, or any potentially dangerous activity. The assistant appears to incorrectly classify normal development tasks as prohibited and refuses to help.

This behavior makes the tool difficult to use for everyday coding workflows because many basic requests are rejected even when they are clearly benign.

Expected Behavior

The assistant should:

• Complete normal programming tasks.
• Answer questions about code.
• Help with debugging and refactoring.
• Generate boilerplate code.
• Perform routine software development actions.

Safety refusals should only occur when the request actually falls into a restricted category.

Actual Behavior

The assistant responds with messages indicating that security policies or safety directives prevent it from helping, even for requests that have no apparent security implications.

As a result, a significant portion of normal coding tasks cannot be completed.

[jatmn]

Please share actual errors and screenshots.
This sounds more related to your selected model or provider.

[Gravirei]

This is cyber risk system remainder when model is not opus 4.6 in fileread tool.

    path: /src/tools/FileReadTool/FileReadTool.ts

[syxs22]

Thanks for finding this, Gravirei.

This is exactly the issue I'm experiencing.