Microsoft Researchers Demonstrate How AI Models Can Be Turned Into “Sleeper Agents”

November 17, 2026


Microsoft researchers demonstrate how tampering with AI model weights can turn AI models into “sleeper agents” that execute hidden malicious instructions when triggered, highlighting a growing AI supply chain security risk.

During a week in which Microsoft grabbed tech headlines for all the wrong reasons, its Microsoft Research division quietly released a report that documents a new potential threat from AI systems. The long and short of it is that all frontier AI models are vulnerable to having some of their neurons responsible for mathematical calculations quietly replaced, the new versions embedded with malicious instructions that are triggered by either a phrase or event like a date change.

The researchers demonstrate that the attack works, but it does involve direct access to model weights to pull off. That is not an easy proposition when it comes to frontier developers, but is much more feasible with smaller platforms and more localized deployments (as the recent Fireworks AI breach demonstrated).

“Lazarus” attacks by AI models possible, but preventable

The attack can potentially be very difficult or very easy, depending on the situation. Someone attempting to breach a frontier developer’s AI models from the outside is going to have a very hard time getting at their model weights; there isn’t any known instance of this yet outside of the compromise of the first version of the Meta Llama LLM in 2023. That incident illustrates the most likely and easiest path to a compromise: a leak or theft by a trusted insider or a partner with privileged access.

And while the frontier developers are hardened, others downstream from them are not as prepared. Fireworks AI is an excellent recent example of a high-profile breach that involved compromise of model weights. There have been a number of similar incidents just in 2026. The attack was also largely developed by Claude, after prompting by researchers to come up with a novel breach method of its own. However, Microsoft stresses there is not yet an indication that this attack has actually been used in the wild as of yet.

Need for Organizational Inventory and Monitoring of AI Stressed Yet Again

AI models now have billions (if not trillions) of neurons that are corruptible given access to the model weights. The research demonstrates it only takes around 10 to execute a malicious command, with most modern models now keeping the vast majority of their neurons dormant while performing their tasks.

That would seem to present an intimidating challenge, but the Microsoft researchers say that frontier AI models are also very good at comparing weights to tell if even very small changes like this have been made. Their prescription is a lot of auditing: auditing before deployment, auditing all third party models, and red teaming before and after each update. Runtime monitoring can also catch malicious commands in the act if they do happen to slip through.

However, it will still be tough for organizations to catch tainted AI models delivered from a trusted source (such as the frontier developers themselves). This appears to be yet another area in which AI developer defenses have to be trusted to hold. But organizations can reduce risk and exposure from other sources with a combination of auditing, monitoring and having emergency responses developed and trained. Most crucially, it is important to have visibility into the history of AI agents with privileged access.