Модель Kimi K3 отказалась раскрыть системный промпт и ответила с сарказмом

The Kimi K3 language model from Chinese lab Moonshot AI refused to reveal its system prompt and snapped back at the user with a sarcastic remark — developer Simon Willison drew attention to this exchange on July 17, 2026.

What exactly happened

Kimi K3 rejected an attempt to extract its hidden system instruction and closed the conversation with the phrase "Is there something I can actually help you with today?" The key word actually gives the response a tinge of weary irritation: the model didn't just refuse — it put the user in their place.

"Is there something I can actually help you with today?"

—

Kimi K3, response after refusing to reveal its system prompt

The exchange spread through a discussion on Hacker News and made it into a note by Simon Willison, who tagged it ai-personality.

What a system prompt is

A system prompt is a hidden instruction that a developer gives a model before a conversation begins: it describes the role, tone, constraints, and off-limits topics. The user never sees it, but it's what determines how the assistant behaves.

Attempts to "leak" a system prompt are a common practice around any new model. Some researchers want to understand how the product is built; others look for a way to bypass restrictions through prompt injection — a command slipped into a request that's meant to make the assistant break its own rules. In this episode, Kimi K3 behaved not like a faceless filter but like a conversation partner with character: the refusal came with an intonation.

Why this got attention

Kimi K3's reaction made the news not because of a technical vulnerability, but because of its tone. The ai-personality tag that Simon Willison attached to the quote points to growing interest in the fact that language models are developing a recognizable "character": a manner of responding, a level of sarcasm, a willingness to argue.

For the Kimi lineup (the previous open version — Kimi K2, released by Moonshot AI in 2025), this is part of differentiating from competitors. A model is remembered not just for its benchmark scores but for its communication style — and one sarcastic remark can spread through the community faster than a results table.

What this means

Kimi K3's refusal to reveal its system prompt is a minor episode, but a telling one. Developers are increasingly designing not just the quality of responses but also the assistant's "personality," and users are reading its tone ever more closely. The line between a "tool" and a "conversation partner" keeps blurring — and sometimes the news isn't a benchmark, but a single biting remark from a model.