← sources

I asked Claude Code to break the tool that hides my secrets from Claude Code

source: reddit (reddit:ClaudeAI)published 2026-11-02 12:29ingested 2026-11-03 02:37

retained text

My MCP server keeps tokens in the OS keychain so the agent never sees them. I told Claude Code: assume you're prompt-injected, you can write any file I can, get my secrets. Every finding needed a failing test before the fix.

The one that stuck: the agent could simply edit the config the next session reads and switch a preset to auto-approve. A `--confirm-auto` flag doesn't help when the agent has a shell.

Lessons for anyone using agents:

* If permissions live in a file, check who can edit that file.
* A `--yes` flag is not human confirmation.
* Read what approval prompts actually show. Mine hid the host.
* If the agent can drive your screen, it can click your dialogs.

document fetched 2026-11-03 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xlvvbm/i_asked_claude_code_to_break_the_tool_that_hides