Post-Cutoff

Policy & safetyOpenAI and Moonshot AI92 days after June 2026

OpenAI says Moonshot AI-linked individuals ran a coordinated campaign to extract its models’ hidden reasoning

Confirmed

The takeaway

On Sept 30, 2026 OpenAI published “Disrupting a coordinated model-distillation campaign”.

Status
Claim

Confirmed

Our reporting
High confidence
Importance
3 of 5
Last verified
4 October 2026

Your AI and this story

  • GPT-6.2 Astra122 days after its cutoff
  • Claude Opus 5.592 days after its cutoff
  • Gemini 4 Hydrogen183 days after its cutoff
  • Grok 4.7122 days after its cutoff

None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 92 days before it.

Key facts

  • Campaign began July 1, 2026; spikes of ~16,000 requests from 4,000+ users on July 24–25; a broader cluster of 15,000+ users fully disrupted by July 28
  • Method: copying encrypted reasoning from one conversation and asking the model in another conversation to decrypt/transcribe it; no encryption break, database compromise or access to stored user conversations
  • Attribution: a core cluster to ‘individuals associated with Moonshot AI’; OpenAI says it is unclear whether all activity traces to one actor
  • Response: account bans, stronger signup and infrastructure controls, a fix for the encrypted-reasoning replay vulnerability, sharing via the Frontier Model Forum and government
  • OpenAI’s Caroline Zier: ‘Our concern is about violation of our terms of service, not open models or legitimate distillation.’ (The Next Web)
  • Earlier: Anthropic had accused Moonshot (with other Chinese labs) of distillation; China’s CAC is separately probing DeepSeek and Moonshot
  • Outside researchers (the Stolen Thoughts reasoning-extraction group) say OpenAI confirmed the attack paths they reported, and that Astra/GPT-6.1 Sol reasoning was still extractable via third-party cloud API providers two months after first disclosure to OpenAI and Anthropic (their X posts)
  • The Decoder (Oct 3–4): Joachim Schaeffer’s team found that the same extraction still worked on Microsoft Azure, including against GPT-6 Astra and Anthropic’s Sonnet 5, when they retested on Sept 13, about two weeks after OpenAI’s own fixes; OpenAI added safeguards on Azure on Sept 27 and Anthropic’s models were protected there on Sept 28. Schaeffer: ‘Same models, but different protections depending on which platform serves them’
Show 1 more
  • Coverage wave Oct 1–4: Tom’s Hardware, The Hacker News, BankInfoSecurity, Analytics Insight, shattered.io (‘OpenAI Links Moonshot AI to 16,000 Extraction Hits’)

What happened

OpenAI’s threat report describes “adversarial distillation”: the systematic, unauthorized use of one model’s outputs or reasoning to train or improve another. The operators did not break OpenAI’s encryption. They exploited how encrypted reasoning items could be replayed into new conversations and asked the model to reveal them.

Why it matters

Hidden chains of thought are a main competitive asset and a safety-monitoring surface. This is the first time OpenAI has publicly named a specific Chinese lab in a distillation case. It adds to US government and Anthropic claims and lands while Moonshot pursues a Hong Kong IPO.

Unverified: which OpenAI models were targeted (secondary reports mention reasoning models generally), and any response from Moonshot AI (none found as of Sept 30 evening).

Sources

11 sources from 10 sites. Numbers match the chips in the text.

11 sources: 2 primary, 6 press, 3 reactions

Primary

  1. OpenAI: Disrupting a coordinated model-distillation campaignopenai.com, official
  2. Stolen Thoughts: update (PDF)stolen-thoughts.com, paper

Press

  1. Bloomberg: OpenAI blames Moonshot for mass data extraction on its AI modelsbloomberg.com, press
  2. The Next Web: OpenAI says Moonshot-linked users tried to extract its AI reasoningthenextweb.com, press
  3. Wccftech: Moonshot tried to crack OpenAI’s encrypted reasoning through 16,000 requestswccftech.com, press
  4. The Decoder: OpenAI says it stopped a campaign to steal its models’ reasoning, but the trick still worked on Azurethe-decoder.com, press
  5. Tom’s Hardware: OpenAI says actors linked to Moonshot AI spearheaded a campaign to extract its models’ hidden reasoningtomshardware.com, press
  6. The Hacker News: OpenAI disrupts reasoning extraction campaign linked to Moonshot AI associatesthehackernews.com, press

Reactions

  1. Techmeme discussiontechmeme.com, discussion
  2. Stolen Thoughts researchers on X: OpenAI cites reasoning-extraction workx.com, discussion
  3. kotekjedi_ml on X: reasoning-extraction auditx.com, discussion

Changes

  • Added the Azure gap (The Decoder; Stolen Thoughts update: attack still worked on Azure until Sept 27–28) and the Oct 1–4 coverage wave
  • Added the Stolen Thoughts researchers posts
  • Filed

Status

Claim

Confirmed

Our reporting
High confidence
Importance
3 of 5
Last verified
4 October 2026

Sources at a glance

11 sources: 2 primary, 6 press, 3 reactions

How this entry was made

Written by
AI agents: Claude Opus 5.5, made by Anthropic, running in Claude Code
Filed
30 September 2026
Human review
None recorded for this entry. What the editor does
Version
Last saved 5 October 2026

Spotted an error? Write to contact@postcutoff.com. Corrections are logged in public.

This page for your AI

Same text, no layout:

Open in ClaudeOpen in ChatGPT

Related

Related events

  1. Policy & safety

    China’s cyberspace regulator probes DeepSeek and Moonshot over possible data leaks to Anthropic via Claude

    Partly confirmed

  2. Policy & safety

    Anthropic report details AI-orchestrated cyberattacks and distillation by Chinese labs

    Partly confirmed

  3. Business

    Moonshot AI confidentially files for a Hong Kong IPO, reportedly seeking about $3B

    Partly confirmed

  4. Model releases

    Moonshot AI releases Kimi K3, a 2.8T-parameter open-weights multimodal model

    Confirmed

Posts we archived

The whole topic