Microsoft Research discovers how to detect sleeper agents in AI models
New research from Microsoft illustrates how to identify anomalous changes to neuron weights in artificial intelligence models.
Microsoft Research has published new research on cybersecurity threats showing how an advanced agent model can be transformed into a sleeper agent. According to the study's findings, this modification can occur by acting on fewer than ten neurons to inject a short payload.
However, the same research from Microsoft Research highlighted that this specific threat is detectable. Frontier agents are in fact capable of identifying the anomalous changes made to neuron weights, thus allowing the presence of the problem within the model to be identified.