cd /news/artificial-intelligence/stolen-valor-how-researchers-discove… · home topics artificial-intelligence article
[ARTICLE · art-112063] src=thestack.technology ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

Stolen valor? How researchers discovered some open-weight models might have cut corners

Research published last week by eight researchers from seven institutions detailed a method for obtaining hidden reasoning traces from proprietary large language models via APIs, potentially enabling model distillation from leading US models. The 'Stolen Thoughts' paper implicated MoonshotAI's Kimi models and Z.ai's GLM-5.2, and author Joachim Schaeffer noted the observations were consistent with distillation, though he cautioned there is no proof.

read2 min views1 publishedAug 26, 2026
Stolen valor? How researchers discovered some open-weight models might have cut corners
Image: Thestack (auto-discovered)

Artificial intelligence is a field within computer science that was built on decades of study, experiments, and collaboration between academia and the private sector. All that work paved the way for everything that came after it, but now that the stakes surrounding AI are the highest they've ever been, that spirit of collaboration has turned into fierce competition that some AI researchers think has gone too far.

Research published last week suggested MoonshotAI, the Chinese company behind the Kimi AI models, may have found a way to “steal” reasoning context from leading US models, and that paper has caught a lot of attention. Z.ai's GLM-5.2 model, which was the model Hugging Face turned to during the OpenAI hacking incident, was also implicated.

Eight researchers from seven institutions put their names to the “Stolen Thoughts” paper, detailing a method for obtaining hidden reasoning traces from proprietary LLMs via APIs by taking encrypted chain-of-thought blocks and replaying them to a weaker model that lacked security protections introduced in newer models.

One of the authors behind the paper, MATS Researcher Joachim Schaeffer, told *The Stack *it would be a big deal if any lab could access the raw reasoning data of frontier LLMs and use it to train their own models.

AI model distillation, the process of using a larger model to generate responses used to train a smaller model with similar capabilities, has already been a point of contention between US and Chinese labs, with Anthropic accusing Moonshot AI, DeepSeek and MiniMax of illicitly deploying the practice on Claude models in February.

“I think one has to put a lot of caveats in here of ‘there's no proof, we just observed interesting model behavior, we only asked questions,’" Schaeffer said. "But nonetheless, the things that we saw were consistent with that [distillation] story."

How it works

Get the full story: Subscribe for free #

Join peers managing over $100 billion in annual IT spend and subscribe to unlock full access to The Stack’s analysis and events.

[Subscribe now](https://www.thestack.technology/membership/)

Already a member? [Sign in](https://www.thestack.technology/signin/)
── more in #artificial-intelligence 4 stories · sorted by recency
── more on @moonshotai 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/stolen-valor-how-res…] indexed:0 read:2min 2026-08-26 ·