ai-devtools —

Malware hides payloads in 0.00001% of open-weight AI models

Malicious code can be concealed within a minuscule fraction of a neural network, specifically within just 0.00001% of it.

Malware hides payloads in 0.00001% of open-weight AI models

This capability raises a critical question for any end-user: how can one detect such embedded threats?

The technique described involves embedding malware within the structure of open-weight AI models, a strategy termed the 'Sleeper Agent' tactic.

This development highlights a growing concern regarding the integrity of foundational AI components, presenting a novel vector for compromise within software infrastructure.

Earlier coverage

Reported by darkreading.com.

ai-devtoolssecurity

Related coverage

Elsewhere in this edition