AI models can hide malware in tiny weight changes, undetectable to users.
Original headline: 'Sleeper Agent' Tactic Hides Payloads in Open-Weight AI Models — darkreading.com
More in Cybersecurity news
Security researchers have identified a new threat: malicious actors can hide malware payloads within the weights of open-source AI models. By manipulating just an infinitesimal fraction of the neural network's parameters—as little as 0.00001%—attackers can embed 'sleeper agent' code that remains dormant and undetectable during standard model validation and performance checks.
For developers and organizations, this creates a critical supply chain risk. Downloading and deploying what appears to be a legitimate, high-performing open-weight model from repositories could introduce a hidden backdoor. The real-world impact is that trust in the open-source AI ecosystem is undermined, forcing a re-evaluation of how we verify the safety of community-shared models. The key takeaway: the very architecture that makes AI powerful also makes it a perfect hiding place for threats.
Read the original at darkreading.com
Satirical commentary on real tech news, summarised with AI. Read the original source for the facts. How ZeroDaily works
