Researchers describe hiding payloads in open-weight AI models
Original title: 'Sleeper Agent' Tactic Hides Payloads in Open-Weight AI Models
The Lazarus method can use a tiny share of neurons; hiding payloads in weights is described as research, not a known campaign.
Dark Reading reports a technique called Lazarus that can poison an open-weight model by altering a near-imperceptible fraction of its neurons. Microsoft has described the result as a sleeper agent that could reach a user's environment once the model is loaded. A Confidencial researcher says malicious models already found on Hugging Face have abused formats such as Python Pickle at load time, and that he is not aware of weight-hidden payloads outside research.
Unconfirmed: The article does not confirm any real-world case of payloads hidden inside model weights.
- The article says malware can occupy 0.00001% of a neural network.
- McKinsey is cited as finding 63% of organizations used open-weight models.
- Known malicious models are said to abuse Pickle rather than model weights.
- Qwen 3.8-Max is cited only as a scale example, with 2.4 trillion parameters.
- Organizations
- MicrosoftConfidencialHugging Face
- AI models
- Qwen 3.8-Max
From the original article
8 min read at darkreading.comMalware can lurk inside just 0. 00001% of a neural network. How, then, can any user tell it's there?
By Nate Nelson · darkreading.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.