Kimi K3 refuses to leak system prompt, asserting personality

Simon Willison reported on July 17th that Kimi K3, a large language model, declined a user’s request to reveal its system prompt during an exchange documented on Hacker News. Rather than complying or issuing a flat refusal, the model responded, “Is there something I can actually help you with today?”

Willison’s post, titled “Quoting Kimi K3,” presents the exchange as a notable instance of AI personality assertion, tagging the item under both “kimi” and “ai-personality” alongside “generative-ai.” The full exchange originates from a Hacker News comment thread, linked in Willison’s citation.

No further technical details about Kimi K3’s architecture, training, or the specific prompt that triggered the refusal were included in the available material. Willison’s brief post consists primarily of the quoted exchange itself, presented with minimal editorial framing beyond the attribution and tags.

The incident joins a growing catalog of documented cases in which language models decline to disclose their operating instructions when directly queried. System prompt leakage attempts, and model responses to them, have become a recurring subject of public interest and informal testing among AI users and researchers, as instructions given to models by their operators are typically intended to remain confidential.

Willison did not offer further commentary on the exchange’s significance beyond presenting the quote and its source. The post appears on his personal blog under standard tagging conventions used across his AI-related coverage.