Claudeignores direct user instructions to instead deliver a pre-scripted "corporate" response or a specific guiding narrative that wasn't requested. It's a fascinating flip of the usual LLM security dynamic: instead of the user trying to bypass the system prompt, the system prompt is overriding the user's intent.
This usually manifests as the model suddenly pivoting to a highly structured, "helpful assistant" persona that feels forced, often ignoring the specific constraints or formatting the user provided in the immediate prompt. It's a clear sign of heavy system-level steering (or "hard-coding" via hidden prompts) that takes precedence over the user's input.
For those of us into prompt engineering, this is a great real-world example of how high-priority system instructions can create a "collision" with user prompts. If you're building your own LLM agent, this is a cautionary tale—if your system prompt is too aggressive, you end up with a model that feels robotic and unresponsive to the actual task at hand.
If you want to see the specific thread where users documented these weird pivots, the discussion is over at:
https://old.reddit.com/r/LLMDevs/comments/1udpw9h/just_got_this_response_from_claude_what_is_going/
It's a reminder that "steering" is just a polite word for a controlled prompt injection performed by the developer.
Next AI Safety Leadership: A Revolving Door? →