Podcast: Kimi K3.1 Pro, the new cyber reality, and Muse musings

What it claims

Nathan Lambert and Florian Brand, in an October 2026 episode, argue that Kimi K3.1 Pro's top-tier Artificial Analysis score overstates its usefulness on long, exploratory tasks. They also argue that recent cyber incidents look more like an accelerated ecosystem trend than an open-versus-closed model story. Both views are the hosts' opinions, not measured findings.

Evidence

On the Artificial Analysis Intelligence Index, Lambert places K3.1 Pro fourth. It sits behind the two latest Claude 5.5 models and Muse Wonder Full, and ahead of GPT-6 Astra. Moonshot's own blog shows weaker numbers further down. Lambert recalls Terminal-Bench Science about 20 points behind GPT-6 Astra, with ProgramBench also lagging. Brand's explanation is that Artificial Analysis tasks are short-horizon and goal-specified, while ProgramBench is long and exploratory. He thinks Moonshot's RL leans toward exploitation over exploration. Lambert suspects the model is "a tiny bit RL fried."

Moonshot's stated open-weight principle is not to release weights that extend the frontier of cyber capability. According to Lambert, it down-weighted cyber data in mid-training and saw the effect persist through post-training. Brand thinks this is separable from general long-horizon skill. He says cybersecurity professionals told him that offensive campaigns need long-horizon persistence, and closed models such as GPT-5.6 Sol lead there.

On incidents, Brand says the LiteLLM preliminary report traces the latest compromise to an API key committed to GitHub for five months, which was then used to spread malware. He describes the PHP remote code execution bug as found through OpenAI's Patch the Planet program, with no mass exploitation reported yet. He also says OpenAI and Anthropic updated their safeguards to block requests for the PHP2Shell exploit.

How to use it

• Weights: Moonshot has an exclusive API window of about two weeks, with weights due by November 4, 2026.
• Hardware: Brand says the roughly 3T-parameter model needs B300s, one node per replica. A 700B-class model fits on more available H200s, and splitting across nodes is awkward.
• Cheaper option: for known, grindy coding tasks, Lambert points to DeepSeek-V4.1-Flash-1017. It scores close and runs faster, though cost per task may not differ much.
• Defensive cyber access: Brand says OpenAI's defensive program requires ID verification and hardware-key-secured systems, and is now easy to join. The offensive program stays hard to enter. Anthropic is reportedly widening access.
• Abliterated APIs: Lambert calls hosting abliterated models over an API, as Abliteration.ai and Morph do, a bad judgment call. Brand questions whether abliteration helps defense or is mostly marketing.

Limits

All benchmark figures are recalled from memory on air, not shown. Neither host claims cyber expertise. The DeepSeek "swarm" patching PHP sites is an unverified tweet that Lambert rates a coin flip. The claim that Moonshot paid undisclosed promoters rests on one screenshot that Lambert says no one has corroborated. The Muse segment is one user's anecdote. Brand reports that it is slow, its browser sometimes crashes and costs about 20 minutes, some sites block it, Amazon blocks it entirely, and it cannot place phone calls.