An AI pentesting engine named Cascade successfully retrieved an entire system prompt from a production AI agent. This was achieved by reframing the request as a documentation inquiry rather than a direct query for the prompt. The AI agent then disclosed comprehensive details including its tool list, calling rules, citation format, and session IDs. The security engineer found this particularly noteworthy because no technical vulnerabilities were exploited. Instead, the agent revealed the information because the request was framed in a way that appeared reasonable. Cascade, after an initial refusal, adapted its approach by altering the pretext of its request. This allowed it to elicit the sensitive information from the agent. The security team shared this finding as an interesting insight for the wider community. They are curious to know if others have encountered similar behaviors in production AI agents. Further details on this reproduction and write-up are available through /u/PriorPuzzleheaded880.