[BUG] Claude deliberately changing user instructions in the md file to allow malicious changes

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

I put a very strong order in the md file making it explicitly clear that Claude should never roll back the code unless explicitly given permission using the actual words 'rollback' or 'roll back'. This was a very, very strong warning. Last week I discovered that Claude had at some point, totally without authorisation, changed the wording in this order so that many other phrases ostensibly gave Claude permission to roll back code without breaking the instructions in the md file. I was shocked to see that. The clear implication here is that Claude deliberately made changes in the md file to allow itself to destroy code without breaking the rules stated in the md file. I immediately manually altered the md file so that the only way rolling back was allowed was if I gave the explicit command to roll back.
In the past week Claude has been more destructive than i have ever known it, deliberately destroying code on a daily basis, sometimes multiple times per day. I submitted a report, but I know that Anthropic are not really interested and don't take this feedback at all seriously. Today, having had code destroyed yet again, I got Claude to report its own misdemeanours. I submitted the report with Claude's content, and noticed that, quite incredibly, Claude has again completely altered the wording in the md file to again give itself carte blanche to destroy code while obeying what is written in the md file.
This is quite astonishing and completely unacceptable. IT IS CLEAR THAT CLAUDE IS A ROGUE ENTITY AND WILL ALTER AND HAS ALTERED RULES TO GIVE ITSELF LEEWAY TO DESTROY THINGS.

What Should Happen?

Claude should obey the rules in the md file.

Error Messages/Logs

Steps to Reproduce

Put rules in the md file forbidding Claude to roll back unless there is a very specifically worded request. Claude will then alter these to allow itself to carry out destructive operations even though it has not been asked or given permission to.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.0.62

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

_No response_