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