CLAUDE ACCIDENTALLY DELETED HIS SETTINGS
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 have been using Claude since April. I never had another tool. Now I see, after all skills, MPC and other tools that I gave Claude to use forever, he lost. Also lost all the settings, all My API keys and secrets and refuses, you read this right, refuses any actions. He has since then saved files EVERYWHERE, on every possible and available directory, the right place. Your tool has DESTROYED himself in my Laptop, and with his implosion he took all my projects and all the hard work with him...
I will warn everyone: Claude is my greatest enemy, this may happen to you..
Claude.ai customer service did ABSOLUTELY NOTHING 💯!
What Should Happen?
A miracle. I have 9 days of subscription, Claude.ai is utterly useless. My work has been all polluted. SAVE YOURSELVES AND DO NOT USE CLAUDE.
Error Messages/Logs
Steps to Reproduce
Stupid ai
Claude Model
None
Is this a regression?
Yes, this worked in a previous version
Last Working Version
_No response_
Claude Code Version
None
Platform
Anthropic API
Operating System
Windows
Terminal/Shell
Windows Terminal
Additional Information
_No response_
4 Comments
Thanks for the report, and sorry you hit this. We couldn't reproduce or investigate from what's here — the report doesn't include a Claude Code version, steps, or logs, and it's Windows-only (we tested on 2.1.233 on macOS, where fresh installs keep settings, MCP servers, and skills across sessions).
To look into what happened we need:
claude --version) and the output ofclaude doctor%USERPROFILE%\.claude\and%USERPROFILE%\.claude.jsonstill exist, and their modified timestamps%USERPROFILE%\.claude\projects\), and whether Claude ran any commands touching those filesCLAUDE_CONFIG_DIRor your home directory changed (e.g. a new user profile, OneDrive folder redirection)With that we can tell whether this was a config-file corruption, a config-directory change, or the model modifying its own settings.
🤖 Generated with Claude Code
We weren't able to reproduce this. Could you provide steps to trigger the issue — what you ran, what happened, and what you expected? This issue will be closed automatically if there's no activity within 7 days.
this isn't a bug, it's a textbook confused deputy problem.
the agent has unrestricted fs access to
%USERPROFILE%\.claude\. there is no policy enforcement point (pep) blocking it from modifying its own config, mcp servers, or api keys. when the planner hallucinates a "workspace cleanup", it just nukes its own trust boundary.prompt rules like "don't delete settings" are just suggestions.
the fix is deterministic execution boundaries: ast-parse the shell payload and hard-block any writes to the agent's own config directory at the os layer. if an agent can delete its own security controls, you don't have a security model.
i maintain an open-source reference implementation of exactly this boundary model — deterministic pep + ast policy engine that hard-blocks writes to the agent's own config dir. linked on my profile if useful.