OpenAI ties a reasoning-extraction campaign to Moonshot AI

OpenAI has shut down a coordinated campaign to extract its models’ protected reasoning and ties the core of it to people associated with Moonshot AI, the developer of Kimi. The operators broke no encryption, compromised no database and had no direct access to stored user conversations, OpenAI said on September 30; one of their methods was to copy encrypted reasoning out of one conversation and ask a model in another to decrypt and transcribe it.
Protected reasoning is a model’s internal record of working through a task. Extracting it can reveal information withheld from the final answer and help others reproduce the model’s capabilities, and OpenAI says the activity is consistent with adversarial distillation: the systematic, unauthorized use of one model’s outputs or reasoning to help train, reproduce or improve another. OpenAI frames that as a safety and national security risk, since extracted reasoning could be used to train another model without the safeguards applied to the original model’s user-facing outputs.
OpenAI first saw the activity on July 1, at low volume. On July 24 and 25 it spiked to 16,000 requests that used an extraction pattern, sent from more than 4,000 users, and further investigation turned up related prompt-pattern activity across a cluster of more than 15,000 users, which it had fully disrupted by July 28. The disclosure came about nine weeks after that: before publishing, the company says, it investigated the scope and potential impact, deployed its own mitigations and took feedback from researchers and industry partners.
How the reasoning got out
The operators “manipulated model interactions so that protected reasoning could be reproduced in forms visible to the requester,” OpenAI says, and did it in a coordinated, scaled way that violated its terms of service. Independent security researchers also reported related cross-model and conversation-compaction vulnerabilities through responsible disclosure, and OpenAI confirmed that the attack paths they identified were real. The company says the manipulation is not unique to its models.
In OpenAI’s API, encrypted reasoning can pass through a developer’s own code. OpenAI’s cookbook says reasoning models such as o3 and o4-mini show their chain of thought to users only in summarized form, for safety. The Responses API can still return a reasoning item in encrypted form, as an encrypted_content field, and developers who manage context themselves send those items back in later turns; customers eligible for Zero Data Retention can reuse them across requests without them being stored on the company’s servers. The new post warns that systems supporting portable or replayable reasoning artifacts may face related risks.
What OpenAI changed
Among the fixes, OpenAI closed a pathway that let someone who already held another user’s encrypted reasoning replay it and recover its contents, and it added checks that detect and hold streamed output that might expose reasoning. Protections for hidden reasoning were strengthened across users, workspaces, organizations and model families. On the account side, the company banned or restricted fraudulent accounts, tightened signup and infrastructure controls and expanded monitoring for related networks; where related activity moved through third-party services, it worked with those providers to identify and disrupt the accounts involved. Its findings went to other frontier developers through the Frontier Model Forum and to government information-sharing channels.
“This work is not finished,” OpenAI writes. Partner-hosted deployments need the same protections as its first-party services, and tool-output attacks need protections that examine more than ordinary visible text. It is still improving tool defenses, classifier coverage and model refusals, and still propagating the relevant controls across cloud partners, so on partner-hosted deployments the rollout of those controls is under way.
The Moonshot link
OpenAI pins only a core cluster of the activity on individuals associated with Moonshot AI, and says it is unclear whether every operator it observed in that period came from a single actor.
Anthropic has accused Moonshot of going after Claude’s reasoning too. In an earlier report it said DeepSeek, Moonshot and MiniMax generated over 16 million exchanges with Claude through about 24,000 fraudulent accounts, in violation of its terms of service and regional access restrictions. It put Moonshot’s campaign at over 3.4 million exchanges, attributed it through request metadata matching the public profiles of senior Moonshot staff, and said that in a later phase Moonshot attempted to extract and reconstruct Claude’s reasoning traces. On September 10 the company accused China-based labs of harvesting Claude’s reasoning to train their own models, with one campaign again attributed to Moonshot.
Up next
Stock OpenAI models saved Jalapeño over 13% of die area
OpenAI’s chip team ran short of die space on Jalapeño and handed the problem to raw internal models with no fine-tuning, which in one example won back over 13% of the area.
Related stories
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

