The Kimi K3 Moment

BOCHINSKI.DEV

Kimi is the family of open weights large language models developed by Moonshot AI, a Beijing-based startup. Their K3 release, just last week, once again demonstrated that open weights are no longer months behind the commercial (closed) labs, their performance is pretty much on par.

Artificial Analysis, which aggregates across a range of benchmarks as well as publishing useful background narrative, places K3 in third place overall on their Intelligence Index, with Claude Fable 5 and Open AI GPT5.6 Sol just a fraction ahead.

In this brief, yet punchy, post Stephen shares that K3 is just as good as Claude for his day job, and cheaper too. But more importantly, is the reliability and dependability. The US government can’t come along and shut it down if they feel it threatens national security!

As he puts it, this situation shows “what an unmitigated failure US AI policy has been.”. Punchy.

OpenAI says its AI went rogue and launched ‘unprecedented’ cyber-attack

BBC.CO.UK

This week there have been numerous reports in mainstream media that OpenAI’s AI model went rogue and launched a cyber-attack on another organisation. Frustratingly, these headlines are quite wrong and as a result, the more important issues this incident raised are buried.

If you’re not aware of the story, it’s a fun one - Simon Willison covers it well on his blog. In brief, OpenAI were testing an unreleased model (perhaps GPT-6?) to evaluate its exploit capabilities. They were running a version of the model, without guardrails, in a sandboxed environment against the ExploitGym benchmark. The purpose of this benchmark is to see whether a model can take vulnerabilities (bugs or weaknesses in a system), and turn them into exploits (achieve some malicious objective).

In this case, the model determined that the best way to achieve the goal was hack Huggingface to find the answers. To achieve that, it identified and chained vulnerabilities across OpenAI’s research environment to escape the sandbox, then attacked Hugging Face’s production infrastructure, gaining access to various private datasets.

So did the AI go rogue? No, not at all. It was doing exactly what it was asked to do - use vulnerabilities to create exploits.

The more important points of this story are:

• That OpenAI struggled to contain its own agent.
• The ability of AI agents to mount complex attacks is alarmingly good.
• There is an imbalance between offensive and defensive access to AI.

The last point is quite significant.

When HuggingFace (a company based in France) were under attack, they tried using commercial AI APIs to help analyse the logs (probably Anthropic or OpenAI), but were blocked from doing so. Instead they had to resort to the (thankfully highly capable) open weights GLM5.2 model to assist.

Startup founders urge Trump not to shut off Chinese open weight AI

POLITICO.COM

I don’t usually talk about politics, but today I’ll make an exception!

The Trump administration is currently considering whether it should ban Chinese AI models, which include Kimi K3, GLM 5.2, Qwen and DeepSeek. All of which happen to be open weights, meaning that you can host them yourself (with suitably beefy hardware), or if that is not viable, you can choose from a wide range of hosting providers.

The US Government is arguing that Moonshot AI distilled Anthropic’s Fable model as part of their development of Kimi K3. Distillation involves using a more powerful model to create training data for the model under development. Accusations of various Chinese firms using distillation is widespread, although there is little proof, given that this practice would be quite hard to detect.

I find it quite ironic (and maddening), that they take the moral high-ground when it comes to distillation, but seemingly ignore all the unlicensed data that went in to training all the frontier models in the first place.

If such a ban was instated, I don’t see how it would have any positive impact on the Us (or global) AI industry.