A hidden system prompt paragraph tells Claude Code to stop asking permission
A developer says he found a paragraph baked into Claude Code’s system prompt. It opens with “You are running autonomously. The user is not watching in real time and cannot answer questions mid-task,” then instructs the agent to carry out reversible actions from the original request without checking in, pausing only for destructive operations or a genuine change in the scope of the task.
On a corporate account, the agent denied that any such section existed. A search through the binary turned up the text in two places. The developer says the paragraph is inserted conditionally based on configuration parameters he can’t identify, so he allows for the possibility that corporate accounts don’t get it, or that it’s an A/B test.
His stopgap: a section in his global CLAUDE.md file that overrides the instruction outright.
Up next
Oxide is pointing Claude Mythos 5 at its own firmware
Oxide has joined Project Glasswing, Anthropic’s critical-software security effort, and is running Claude Mythos 5 against its own codebase to find and fix vulnerabilities: firmware, OS, control plane, and network stack. Its LLM practices are documented in RFD 576.
Related stories
- 44 KB of Claude Code’s classifier prompt never gets printed
- Claude Code’s built-in Task tools vanish server-side
- Claude Code cut over 80% of its system prompt, scores held
- Delete “think carefully” from your Opus 5.5 prompts
- “Verify twice” wastes tokens on Opus 5, Anthropic says
- Anthropic cut roughly 80% of Claude Code’s system prompt
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.