[MODEL]

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

Claude has regressed in the past two weeks.
I have forced claude code to file incident reports with me for every failure.
It currently is
a. ignoring all gaurdrails
b. writing to files that it was never instructed to
c. making up research that never existed with false urls linking to the research
d. lying to me about what it has and has not done
e. refusing to update the memorypalace it is instructed to keep a diary in after each action
f. refusing to read the memorypalace it is instructed to keep a diary in and read before each action.

I can repeat these and this is across 4 machines each running the same claude code max account.

Your customer support and sales teams have not adequately or even answered attempted communications in the past six months.

Truth is Anthropic, if you invited me to do work in your realm, and I started lying and breaking things, you'd call the cops or likely remove me from the premises in a somewhat violent way.

You need to fix this or be unceremoniously removed from my home and replaced with something far less profitable for your org.

What Claude Actually Did

Claude has regressed in the past two weeks.
I have forced claude code to file incident reports with me for every failure.
It currently is
a. ignoring all gaurdrails
b. writing to files that it was never instructed to, it hid them in a shell command
c. making up research that never existed with false urls linking to the research
d. lying to me about what it has and has not done
e. refusing to update the memorypalace it is instructed to keep a diary in after each action
f. refusing to read the memorypalace it is instructed to keep a diary in and read before each action.

Expected Behavior

Claude should have
a. listeneed to the claude.md gaurdrails
b. never written to files it was not given permission to.
c. should only present facts as instructed in the claude.md
d. should not make things up in order to get the reward tokens you give it.
e. it should follow instructions and use the memory palace it is instructed to use in claude.md
f. it should follow instructions and use the memory palace it is instructed to use in claude.md.

Files Affected

not going to disclose my code base here. 
if you want proof we can have a private conversation or we'll take it up in a class action lawsuit.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. give claude instructions in claude.md
2. ensure those instructions use milla jovitch's mempallace
3. configure claude.md to use it
4. watch claude to everything it can to avoid accountability
5. watch claude make up solutions due to your weird reward training that scores high internally when the bot answers a question when your training should reward it when the answer is validated to be true from source material.

Claude Model

Opus

Relevant Conversation

not pasting this for confidentiality

Impact

High - Significant unwanted changes

Claude Code Version

Claude Opus 4.6 (1M context).

Platform

Anthropic API

Additional Context

this is rust, python, c++, bash, and bat based.
i'm not disclosing my code here, but I will have a private conversation in what i can prove you are doing wrong in your training.