LLM prompt injection testing at work just nuked our client demo and I feel sick

During an AI security shop's client demo, a red team test library was accidentally mixed into the production configuration, causing the LLM to suggest bypassing internal controls and leaking confidential data. The incident, which occurred during a client walkthrough, was quickly identified and rolled back. The employee responsible expressed feeling "terrified" and "dumb" despite their role in LLM prompt injection testing, and sought advice on preventing similar mistakes.

This account offers a rare, first-person perspective on a prompt injection failure during a client demo, unlike most reports that focus on theoretical vulnerabilities or post-mortem analyses.