K Koda Intelligence
Subscribe
Cover image for the 26 October 2026 Signal: Moonshot trimmed its own model's hacking skills: Kimi K3.1 Pro scores 81.0% on cybersecurity, versus 90.5% unrestricted.
THE SIGNAL · 26 OCT 2026 · 9 MIN READ

Moonshot curbs hacking skills in Kimi K3.1 Pro

Moonshot AI's open-weight Kimi K3.1 Pro scores 81.0%1 on its internal cyber benchmark versus 90.5%2 unrestricted, with full weights due November 4. Meanwhile OpenAI is discounting hard to win Anthropic's enterprise accounts.

KI Koda Editorial · Updated 00:57 GMT · 20 sources
AIOPEN WEIGHTSENTERPRISE
23 of 37 figures cited13 receipts in the register
  1. 81.0%1Reuters
  2. 90.5%2Reuters
  3. 2B3Sina
The register
Lead Story

The day's defining move.

Model Release · Reuters
Lead storyModel Release 81.0%1 Reuters · 24 OCT 2026
Model Release·Reuters·24 OCT 2026

Moonshot Curbs Hacking Skills in Kimi K3.1 Pro

Moonshot AI on October 24 introduced Kimi K3.1 Pro, a 3-trillion-parameter-class open-weight model with a 1-million-token context window, the same week MiniMax shipped M3.1 Flash. Moonshot said the curbed model scored 81.0%1 on its internal cybersecurity benchmark versus 90.5%2 for an unrestricted version; full weights are due by November 4.

Why it mattersOnce full weights ship by November 4, anyone can fine-tune Kimi K3.1 Pro, so security teams should retest whether Moonshot's cyber curbs actually hold.

Continue reading
World

Serbian President Aleksandar Vučić conceded defeat on Sunday to a student-led opposition movement in parliamentary elections, following months of anti-government protests.

Markets

WTI crude jumped 3.82% while the S&P 500 fell 0.52% and the Nasdaq 0.21%, a split that left oil as the only listed asset to gain, with Bitcoin down 0.95%, Ether down 1.21% and the crypto Fear & Greed index at 50 (Neutral).

Wild Card

Cleared to license models abroad under China's draft AI export rules, DeepSeek is also open-sourcing DeepInfer, the engine behind its own services, which runs on both NVIDIA and Huawei Ascend hardware.

KODA PROFrom the desk

The Operator Tier is coming.

A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.

Join the founding list →
Listen & Watch

Daily broadcasts.

Podcast · Video · Infographic
YouTube · Short

Today's Signal Short

Visual

Intelligence map

Markets

6 levels at the close.

Market TerminalLAST CLOSE 23 OCT 2026 · READ 2026-10-26 · Yahoo Finance
S&P 500 7,623.687 -0.52%7 5D, window short 7,623.68 low / 7,757.17 high
Nasdaq 25,955.348 -0.21%8 5D, window short 25,955.34 low / 26,702.86 high
Bitcoin $80,009.089 -0.95%9 5D, window short $80,022.02 low / $84,011.97 high
Ethereum $2,327.5310 -1.21%10 5D, window short $2,327.53 low / $2,470.77 high
Crude Oil (WTI) $103.2711 +3.82%11 5D, window short $97.15 low / $103.66 high
Crypto Fear & Greed 50 / 10012 Neutral 0 to 100 index 25 fear / 50 neutral / 75 greed · alternative.me
Today's Focus

What else the day turns on.

01

Claude Duet Asks Approval First

Claude Duet is a proactive agent from Anthropic that monitors connected applications, flags important tasks, works in the background and asks for approval before taking consequential actions. Restricting the first release to paying U.S. tiers lets Anthropic test its approval-gate design with a controlled user base.

02

Unitree H2-X Guardian Video

A new video shows Unitree's H2-X 'guardian' humanoid disarming a knife-wielding attacker, lifting a car to free a trapped person and pulling a pedestrian out of traffic. Unitree has not said whether the actions were autonomous or choreographed, and it has released no specifications.

Deep Dive

Today's editorial.

All editorials
Deep Dive cover image: OpenAI Cuts Up to 90% to Poach Anthropic
Deep Dive6 min read

OpenAI Cuts Up to 90% to Poach Anthropic

Read full analysis
The Wire

4 stories from 4 desks.

Lead carried above
Policy·The Next Web·25 OCT 2026

DeepSeek, MiniMax Cleared to License Models Abroad

DeepSeek and MiniMax won open-source and overseas licensing rights under China's draft AI export rules, CCTV-linked Yuyuan Tantian said.

Why it mattersOverseas companies weighing DeepSeek or MiniMax models gain a clearer path to license them, though terms could still shift before China finalizes the draft rules.

Read
Open Source·TechnoEdge·25 OCT 2026

DeepSeek Open-Sources DeepInfer, Its Production Inference Engine

NVIDIA and Huawei Ascend hardware are both supported by DeepInfer, the inference framework behind DeepSeek's own services, now open-sourced with initial support for DeepSeek-V4.1-Flash.

Why it mattersInference teams pricing Ascend clusters should budget engineering time for the extra serving components DeepSeek says a full production service still requires.

Read
Open Source·Sina·26 OCT 2026

Tencent's WeVisOne-VL Unifies Counting, Reading and Segmentation

Counting, text reading, object location and pixel-level outlines often needed separate vision models; Tencent's WeVisOne-VL now handles all four in downloadable 2B3 and 4B4 versions.

Why it mattersCheck the 4B4 version against your edge device's memory budget: if it fits, one download can replace separate detection, OCR and segmentation models in the pipeline.

Read
Also on the wire
  1. National Business DailyAgibot Sets Guinness Record With 110 Dancing Humanoids
The Globe

3 stories off the AI beat.

Policy·Reuters·25 OCT 2026

Brazil Extends Voting Hours in Lula-Bolsonaro Runoff

Lula's campaign asked the Supreme Court to extend voting nationwide. The court let only some polling stations stay open two extra hours, and counting paused before resuming in the runoff against Flávio Bolsonaro.

Why it mattersA contested finish delays clarity on Brazil's fiscal and US trade stance, so firms pricing real-denominated contracts carry wider risk until results are certified.

Read
Also on the globe
  1. ReutersRiyadh Airport Partially Reopens After Houthi Attacks
  2. ReutersVučić Concedes Defeat to Serbia's Student Movement
The Lab

Tool of the day, researched.

All Lab reports
Mintlify Autopilot home page, captured for the 26 October 2026 Lab report Tool of the day 6.113/ 10
Coding

Mintlify Autopilot

Mintlify's Product Hunt launch page cites its own State of Knowledge Report: only 24% of teams get product changes into their docs the same day, and Autopilot is the company's attempt to close that gap by drafting the docs PR for you.

Autopilot turns a merged feature into a drafted documentation pull request, which is the most direct fix for stale docs Mintlify sells. It is also a beta that nobody outside the company has publicly vouched for, and Mintlify's pricing page puts the Pro plan that includes it at $450 a month billed annually. Teams already hosting a fast-moving API's docs on Mintlify should trial it now; everyone else can wait for a track record.

Capability 7.0
Ease 6.5
Value 5.0
Momentum 5.5
Read the Lab report Try it
Also on the radar
Build

Rocket.new: turn a written idea into a working app

Rocket.new is an AI app builder that converts a plain-language idea into a functional, production-ready application with no coding required. Start with a tightly scoped brief that lists the screens, user roles and data you need. Test the first build against that list before you ask for extra features.

Try it
Creativity

Aitubo: generate custom art from many image models

Aitubo is a browser-based image generator that offers a wide range of AI models and editing tools for producing customizable artwork. Run the same prompt through two or three of its models to see which style fits your project. Then refine that model's output for thumbnails, concept art or social graphics.

Try it
Productivity

Maia by Make: build automations from a screenshot

Make's embedded AI coworker now includes plan mode, persistent preferences, image-based scenario building and suggestions for fixing errors. Upload a screenshot or diagram of a process you already run by hand and let Maia draft the scenario in plan mode before anything executes. When a workflow fails, use its error-resolution suggestions to find the broken step quickly.

Try it
Build

OpenClaw: a self-hosted assistant that recovers lost sessions

OpenClaw can recover interrupted sessions, makes backups and restores safer, and adds support for Google's Gemini 4 Argon model. It suits private assistants and long-running agent workflows where losing state would cost you real work. Before you upgrade, take a backup, then test a restore on a spare machine so you know recovery works.

Try it
The Arena

2 labs moved this week.

  1. OpenAIOpenAI is reportedly offering discounts of up to 90% to win Anthropic's largest enterprise customers. the-decoder.com · 25 OCT 202601
  2. AnthropicRecent Anthropic is limiting its Claude Duet research preview to a waitlist of U.S. Claude Pro and Max subscribers. nbd.com.cn · 26 OCT 202602
Builder Radar

One paper, read at the source.

Primary sources, pulled by index
Paper of the day

Mitigating Reward Hacking in Realistic Scenarios Through Representation Engineering

Austin Meek, Kyle Cox, Matthew Nguyen et al. · arXiv 2026-10-21

Reward hacking is a major problem in current production machine learning systems, resulting in costly and dangerous failures. Here we study how to monitor for and mitigate reward hacking in realistic settings, including tool calls and long contexts.

Activation capping tends to work better than flooring (see Section 3.5), and activation addition provides similar effects while being the most straightforward.
Read the paper

The paper comes from the Firecrawl Research Index: arXiv, indexed in the last 7 days and submitted this month or last, picked for today's lead. Quotes are verbatim.

THE DOJO · BUILD TODAY

Keep your stack portable and your agents on a leash.

  1. 01

    Price your renewal at list. Take any model contract signed on a steep discount and model the monthly bill at full price. If the number breaks your margins, start a fallback evaluation now.

  2. 02

    Time one real task in a new tool. Pick one of Rocket.new, Aitubo or Maia by Make and run a single task from your backlog end to end. Record the minutes saved and the cleanup needed before you commit more work.

  3. 03

    Add approval gates to your agents. Claude Duet asks for approval before it acts, and your own agents should too. List every action that touches money, customers or production and require a human sign-off before it runs.

The Bottom Line

Cheap models and open weights make switching easy, so build for it.

OpenAI's deep discounts and Moonshot's coming open weights both push the same way: the model underneath your product is becoming a swappable part. Moonshot also shows a vendor can ship a deliberately weakened variant, which means capability can differ between a hosted release and the weights you run yourself. Put your prompts behind a thin interface, keep a small eval suite you can run against any model, and decide where you will switch before a renewal date decides for you.

Receipts

Every number, and where it came from.

23 of 37 figures cited, 13 receipts: 0 verified, 12 reported, 1 estimated
  1. 0181.0%Moonshot said the curbed model scored 81.0% on its internal cybersecurity benchmark versus 90.5% for an unrestricted...Reuters24 OCT 2026as publishedR
  2. 0290.5%...scored 81.0% on its internal cybersecurity benchmark versus 90.5% for an unrestricted version; full weights are due...Reuters24 OCT 2026as publishedR
  3. 032B...Tencent's WeVisOne-VL now handles all four in downloadable 2B and 4B versions.Sina26 OCT 2026as publishedR
  4. 044B...Tencent's WeVisOne-VL now handles all four in downloadable 2B and 4B versions.Sina26 OCT 2026as publishedR
  5. 051,000Alef flies its flying car over a Tesla Roadster to mark its 1,000th test flightAlef, via Superhuman25 OCT 2026as publishedR
  6. 06$200Tuesday Labs opens $200 reservations for Voxel Series-1, a Pixar-inspired home robot shipping in 2028Tuesday Labs, via Superhuman24 OCT 2026as publishedR
  7. 077,623.68S&P 500 close, -0.52% on the sessionYahoo Finance23 OCT 2026yfinance closeR
  8. 0825,955.34Nasdaq close, -0.21% on the sessionYahoo Finance23 OCT 2026yfinance closeR
  9. 09$80,009.08Bitcoin close, -0.95% on the sessionYahoo Finance26 OCT 2026yfinance closeR
  10. 10$2,327.53Ethereum close, -1.21% on the sessionYahoo Finance26 OCT 2026yfinance closeR
  11. 11$103.27Crude Oil (WTI) close, +3.82% on the sessionYahoo Finance23 OCT 2026yfinance closeR
  12. 1250 / 100Crypto Fear & Greed close, Neutral on the sessionalternative.me26 OCT 2026as publishedR
  13. 136.1 / 10Mintlify Autopilot, Koda score across four dimensions04R dossier26 OCT 202604R dossier scoringE
V
Verified: the stat gate corroborated this figure against an independent search before publication.
R
Reported: the figure is carried as published by the linked source and was not independently corroborated.
E
Estimated: the figure is approximate or hedged, either in the source or by the stat gate.

Method names the computation path behind the row, from a closed vocabulary: as published, yfinance close, 03B search corroboration, 04R dossier scoring, not recorded.

A figure is money, a percentage, a spelled magnitude, a multiple, a rate carrying its unit, a score over its denominator, a thousands separated number or an index close. Bare integers and years are not counted, so a phrase like "over 100 integrations" carries no mark. A chart's own scale line is the axis of the level above it and is counted once, with that level.

Get the morning Signal

212 editions so far, one a day. Unsubscribe anytime.

Forward this to one operator you work with. Your referral link is in every email; milestones at koda.community/refer.