OpenAI says it cut off a distillation campaign tied to Moonshot AI
OpenAI says it disrupted a coordinated campaign to extract protected reasoning from its models. It attributes a core cluster of the activity to people associated with Moonshot AI, the developer of Kimi. The earliest activity began on July 1.

OpenAI published its account of the campaign on its website on 4 October. The company says the operators never broke its encryption, compromised a database or gained direct access to stored user conversations. Instead, they pushed model interactions into forms that reproduced protected reasoning in visible output, at scale, in breach of its terms of service.
That distinction matters for anyone building security tooling on top of frontier models. The attack surface was the product's own reasoning trace, not a server room. OpenAI describes the activity as consistent with adversarial distillation. That means the systematic use of one model's outputs to train, reproduce or improve another.
The company's timeline is unusually specific. Activity started on 1 July at low volume. High-volume spikes followed on 24 and 25 July, with 16,000 requests using a relevant extraction pattern from more than 4,000 users. A wider cluster of more than 15,000 users running related prompt patterns was fully disrupted by 28 July, OpenAI says.
One technique stands out. OpenAI says operators copied encrypted reasoning from one conversation and asked a model in another conversation to decrypt and transcribe the hidden content. If that works, the practical barrier between a model's final answer and its internal working is thinner than customers might assume. OpenAI says it has since deployed mitigations and shared details with industry partners through the Frontier Model Forum.
Why a reasoning trace is worth stealing
Protected reasoning is the model's internal record of working through a task. OpenAI argues that extracting it can reveal information withheld from the final answer. That is exactly the material a competitor would need to rebuild capabilities. The company also warns that a distilled model may not carry over the safeguards applied to the original's user-facing output. At scale, it says, distillation accelerates capability transfer without matching investment in safety.
OpenAI says it is unclear whether every operator observed in the period came from a single actor. It attributes only a core cluster to people associated with Moonshot AI. Moonshot has not responded publicly in OpenAI's post, and the post does not say whether OpenAI contacted the company before publishing. Neither side has disclosed any legal step.
The disclosure follows a pattern in which labs publish incident write-ups only after mitigations are in place. OpenAI says it investigated scope, deployed fixes and took feedback from researchers and industry partners before going public. Independent security researchers also reported related cross-model and conversation-compaction vulnerabilities through responsible disclosure. OpenAI says it confirmed those attack paths were real.
For buyers of AI security products, the takeaway is narrower than the headline. The incident was not a classic breach with a stolen credential. It was abuse of a legitimate interface, repeated enough times to look like an attack.
Google redraws the model menu under users' feet
Model access is shifting at the same time. According to an updated Google support document reported by 9to5Google on 3 October, Gemini app users without a paid subscription will be limited to the 3.5 Flash-Lite model starting 9 October. They lose access to 3.6 Flash and 3.1 Pro.
Subscribers to AI Plus at $4.99 a month will be limited to Flash-Lite and Flash, with Pro removed. Google says those users will receive an email explaining when the changes take effect. AI Pro at $19.99 a month and AI Ultra at $99.99 or $199.99 a month keep access to all three models. AI Pro customers also gain the Deep Think option for maximum parallel reasoning, previously confined to the Ultra plans.
THE DECODER reported on 4 October that the setup mirrors OpenAI's freemium structure for ChatGPT, with the free tier reduced to the smallest model. The publication also noted that Google may be preparing for Gemini 4 Argon, which could be more expensive to run. 9to5Google reported that Argon is first coming to AI Ultra subscribers. It also reported that Google has not said whether Argon counts as a Pro model or sits in a tier above it.
The change has a security dimension that is easy to miss. Teams that built internal tools on whatever model the chat box happened to serve may find themselves on Flash-Lite after 9 October. Prompts tuned for Pro will behave differently, and nobody sends a migration notice for a default.
Google is also preparing low, medium and high effort levels for each available model in October, according to 9to5Google. Higher effort produces more thorough answers and consumes more of a user's limit. That is a billing detail as much as a product one.
Anthropic asks first, and keeps the switch off
Anthropic has begun asking Claude users to volunteer voice conversations for model training, BleepingComputer reported on 4 October. The prompt appears when people use Claude's voice features. It asks permission to use the audio to improve the company's models.
The setting is off by default, according to the interface BleepingComputer saw. Users can pick Allow or Not now, change the toggle later under Settings, Privacy, and delete the voice data afterwards. Voice training is separate from the existing option to train on chats and coding sessions, so the two choices can be set independently.
That separation is the interesting part. A user can hand over voice recordings while keeping ordinary chats and Claude Code sessions out of training, or the reverse. Anthropic has not said how many people have opted in, and the report gives no figure for how much voice data it holds.
Small models, local machines, and a CVE audit
Security teams that cannot send data to a hosted API now have more local options. Jerry Gamblin, who serves on the CVE Program's Quality, Automation and Consumer working groups, ran Cloudflare's Clef models locally through Ollama 0.35.1. He asked two questions of all 27,489 CVEs published in the 60 days to 3 October: does the description state a security impact, and how clear is it for a defender?
The results, published on his blog on 4 October, put numbers on a familiar complaint. Outside the Linux kernel, 989 of 23,736 descriptions, or 4.2%, never say what an attacker could do. The kernel is its own case: its published policy leaves impact to the user, so 3,661 of its 3,753 records state none. That takes the all-CVE figure to 4,650 of 27,489, or 16.9%.
Clef Flash, a 9B model, answered both questions at a median of under a second per CVE on a laptop, with no API key. Flash proved too strict on its own, so the larger 27B Clef model re-read everything it flagged outside the kernel and overturned 1,198 of 2,187, or 54.8%. Against 129 CVEs Gamblin labeled by hand, the two-step approach caught 15 of the 16 he judged as stating no impact and never flagged one of his positives. Both models are Apache 2.0 and fine-tuned from Qwen models, he writes.
Gamblin states that nothing in his post is a CVE Program position. The audit is one researcher's weekend project, not a regulator's finding. It measures prose quality rather than whether a bug is exploitable.
OpenAI's own post makes a related point: distillation and reasoning extraction are not vulnerabilities unique to its models. The company says similar techniques may affect other advanced systems, which makes coordination across labs part of the defence.
Sources
9- 01Disrupting a coordinated model-distillation campaignEN
- 02Gemini app limiting what models free and AI Plus users can accessEN
- 03Google's new Gemini tiers cut free users to its weakest model and lock $5/month subscribers out of ProEN
- 04Anthropic asks Claude users to share voice data for AI model trainingEN
- 05Decision Models vs. CVE DataEN
- 06OpenAI's latest features take direct aim at the app store modelEN
- 07Ideogram 4.5: The most precise edit modelEN
- 08No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning CollapseEN
- 09Language Model "Shape"EN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.