Skip to content
EVENT / TOPICHistorical

Vitalik Buterin supports Moonshot AI's reduction of Kimi K3.1 Pro hacking capabilities

2 reports2 sources10/29/2026 updated

Current overview (not historical)

AI overview

Vitalik Buterin explicitly endorses Moonshot AI's decision to deliberately reduce hacking skills in the Kimi K3.1 Pro model, arguing that limiting AI capabilities causes less harm than discriminating against users. According to Moonshot, the filtered model solved 81% of tasks on its internal exploit benchmark, while an unfiltered version solved 90.5%, though general coding tasks remained included. A follow-up report indicates that Moonshot AI, along with DeepSeek, uses OpenAI GPT-6 Astra outputs to train models like Kimi K3.1 Pro, though this training context does not alter the core decision on capability reduction.

Written with AI from the reporting · updated 10/29/2026

Latest developmentOct 29, 21:20
Chinese AI labs shift to OpenAI's GPT-6 Astra after Anthropic tightens defenses
Explore the timeline, relations and evidence
Expand the detailed fact route and sources

Story route

A station-by-station route from the first public evidence onward; every station carries its own evidence and uncertainty.

  1. 1Unclassified fact stationOct 26, 00:00recorded Oct 26, 15:44single source

    Vitalik Buterin backs Moonshot AI's Kimi K3.1 Pro hacking sk

    Read original source ↗ BeInCryptoReplay this moment in the explorer
    Evidence & revisions · 1 source records
    Subject：
    Moonshot
    Action：
    deliberately cut hacking skills from Kimi K3.1 Pro model
    Object：
    Kimi K3.1 Pro model
    Occurred (fact side)：
    Oct 26, 00:00
    BeInCrypto ↗Report pageOriginal / first version
    Published (valid side)：
    Oct 26, 15:33
    First seen (known side)：
    Oct 26, 15:44
    Fact stationSingle source · one-sided claim

    Copy may have been revised later; time replay is bounded. Layer labels group reading only, not causation between layers.

  2. 2Unclassified fact stationOct 29, 00:00recorded Oct 29, 21:28single sourcerevised

    OpenAI GPT-6 Astra 被中国 AI 实验室作为主要训练目标

    Read original source ↗ The DecoderReplay this moment in the explorer
    Evidence & revisions · 1 source records
    Subject：
    openai
    Action：
    成为主要训练目标
    Object：
    GPT-6 Astra
    Occurred (fact side)：
    Oct 29, 00:00
    The Decoder ↗Report pageRevision 2 · currently revised
    Published (valid side)：
    Oct 29, 21:20
    First seen (known side)：
    Oct 29, 21:28
    Fact stationSingle source · one-sided claimRevised

    Copy may have been revised later; time replay is bounded. Layer labels group reading only, not causation between layers.

Uncertainty marks flag where to check the sources; they are not proven causation.

Report timeline

Explore every side through dated source reports.

Oct 29
  1. The DecoderRelated fact: single sourceSource report revised
    Chinese AI labs shift to OpenAI's GPT-6 Astra after Anthropic tightens defenses

    Report first recorded Oct 29, 21:28

    Chinese AI labs including DeepSeek and Moonshot AI are primarily using OpenAI's GPT-6 Astra to train their latest models, including Kimi K3.1 Pro and V4.1-Flash 1017, after Anthropic tightened Claude's defenses. These labs rely on extracting the models' hidden reasoning steps, a process known as distillation, which Anthropic says has reduced data leakage by 90 to 99 percent.

    Read original source ↗
Oct 26
  1. BeInCryptoRelated fact: single source
    Vitalik Buterin backs Moonshot AI's decision to strip hacking skills from Kimi K3.1 Pro model

    Report first recorded Oct 26, 15:44

    Ethereum co-founder Vitalik Buterin supports Moonshot AI's decision to deliberately remove top-end hacking capabilities from its Kimi K3.1 Pro model, arguing that limiting AI functionality is less harmful than discriminating against users. Moonshot reports the filtered model solved 81% of tasks on its internal exploit benchmark, compared to 90.5% for the unfiltered version. Buterin warns that if the first wave of AI risk goes badly, access could become restricted through KYC checks and surveillance, creating tiered user classes. He also notes that rival models like DeepSeek's V4.1-Flash already achieve 85.7% on the same benchmark, suggesting the capability gap may not be permanent.

    Read original source ↗

Event activity trend

Latest comparable
3
Comparable peak
10Oct 29, 22:00
24h change
-50%

The chart compares consistently observed participants only. Hover or tap to inspect each hour; use arrow keys to move between observations.

Event discussion

Reader opinions are hosted on GitHub Discussions; posting requires a GitHub account. Comments are not verified facts.

Open GitHub Discussions