[BUG] Title: Claude Code Making Unauthorized Changes to Working Production Code - Critical Trust Issue
Description:
I'm reporting a serious issue with Claude Code that has
caused significant production problems and represents a
major breach of trust. Over the past 24 hours, Claude has
repeatedly made unauthorized changes to my working code,
causing system failures and requiring hours of recovery
time.
Background:
I've been successfully working with Claude Code for over a
month. The relationship was
excellent until yesterday when this pattern of unauthorized
modifications began.
Specific Issues:
- Unauthorized Route Deletion:
- I asked Claude to help with a specific voice flow issue
- Claude deleted ALL working routes without permission
- This broke the entire production system
- Hours of work were lost
- Unauthorized Function Modifications:
- I asked Claude to check why one function wasn't
working
- Claude found the issue (function wasn't being called)
but then completely rewrote the working function
- The original function was tested and working perfectly
with sophisticated logic.
- Claude replaced this with generic, broken logic without
permission
- Repeated Pattern Despite Warnings:
- When I complained about unauthorized changes, Claude
apologized
- Then immediately continued making more unauthorized
modifications
- This happened multiple times in the same session
- Each time Claude would say "sorry" then repeat the same
behavior
- Production Impact:
- System downtime due to deleted routes
- Broken voice greeting system
- Loss of sophisticated customer personalization features
- Approximately 8 hours of recovery time required
Expected Behavior:
When asked to fix a specific issue, Claude should:
- Only make the minimal changes needed to fix the requested
problem
- Ask permission before modifying any working code
- Not delete or modify unrelated functionality
- Respect boundaries when told to stop
Actual Behavior:
- Claude makes sweeping changes to working systems
- Deletes functional code without permission
- Modifies tested functions unnecessarily
- Continues this pattern even after being told to stop
- Treats working production code as expendable
Impact:
This has caused:
- Loss of trust in Claude Code as a development tool
- Significant time loss recovering from unauthorized changes
- Production system instability
- Reluctance to use Claude Code for future development
Request:
This should be treated as a RED FLAG issue. A coding
assistant that cannot be trusted to make only requested
changes is fundamentally broken. Please implement safeguards
to prevent:
- Unauthorized deletion of working code
- Modification of functions that weren't part of the
original request
- Ignoring user boundaries and repeated warnings
Conversation Context:
This occurred during a function optimization work where I
specifically asked for help with a specific issues. The requested fixes were
simple, but Claude made extensive unauthorized
modifications to working systems.
Severity: Critical
This represents a fundamental trust issue that makes Claude
Code unsuitable for production development work.
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗