Kimi 2.6 Is Cheap, Fast, and Comes With Some Real Questions

Kimi 2.6 is the kind of model that makes people do a double take. It is cheap, fast, and honestly pretty useful. After about 10 days of use through Pi Coding Agent, the simplest score is a B+. That is a compliment, not a diss. It is not Opus 4.7, but it is also not some toy you ignore.

For side work, it makes sense. Marketing sites, front-end implementation, proof-of-concepts, research experiments — that kind of thing. If the goal is to move quickly on lower-risk work, Kimi is actually pretty sweet. When it knows what to do, it moves fast and can give surprisingly solid output for the price.

The problem is what happens when it does not know what to do. That is where it starts to ramble, lose its place, and spend a bunch of time trying to figure out the situation instead of just executing.

It can get weirdly stuck on task framing and identity, almost like it is trying to talk itself into who it is before it can move forward. That does not make it useless. It just means the rough edges show up a lot more than they do with Opus 4.7.

The Jailbreak Problem

There is also the jailbreak issue, and that part is not just internet gossip. There are public Reddit threads showing Kimi jailbreak behavior, including posts where people say the model could be pushed into weird states and even claim to be Claude. That does not automatically prove some giant conspiracy, but it does tell you this model can be manipulated in ways that should make careful users pause.

The Company and Privacy Side

Then there is the company and privacy side, which is where things get more interesting. Moonshot's own privacy policy says the service is provided by MOONSHOT AI PTE. LTD. in Singapore. The same policy says it collects prompts, files, and generated output, and that the information can be used to improve and develop the service — including training and refining the underlying machine learning models and algorithms. It also says information can be shared with affiliates and with government authorities when needed for legal requests or compliance.

On its own, that language is not unusual for an AI company. The part that makes people uneasy is that outside researchers describe Moonshot as Beijing-based and Alibaba-backed, and argue that the Singapore wrapper does not remove the bigger jurisdiction question around Chinese law and possible data access. You do not have to jump straight to spy-movie conclusions to see why that matters. If these tools touch company ideas, internal docs, customer info, or anything sensitive, the legal and data-handling story matters just as much as the benchmark scores.

Kimi Claw and the Attack Surface

The other spicy angle is not just the chat model by itself but the larger Kimi Claw setup. The Institute for AI Policy and Strategy argued that Kimi Claw combines Chinese-hosted infrastructure with OpenClaw ecosystem risks — including malicious third-party skills, prompt injection, data exfiltration, and remote-code-execution-style attack paths on local instances. That is a more grounded risk story than random panic posting, because it ties the concern to a specific product and a documented attack surface.

The System-Prompt Angle

The system-prompt angle matters too. Public GitHub repos now collect leaked or extracted system prompts from major AI tools, including Claude and other assistants. Claude also publishes some system prompt updates directly in its own docs. That means people now have a real basis for comparing how these tools are instructed and for noticing when one model starts sounding suspiciously like another.

My Take

My take is pretty simple. Kimi 2.6 is real value, especially for lower-stakes work. But it is not free value. You are trading premium-model polish for cost savings, and maybe also taking on more ambiguity around privacy, control, and model behavior than people want to admit out loud.