Georgia Tech researchers uncover AI backdoor threat to self-driving cars

Got story updates? Submit your updates here. ›

Researchers at Georgia Tech have discovered a new vulnerability called VillainNet that exposes a critical blind spot in autonomous vehicle systems. VillainNet is a dormant AI backdoor that can be programmed to activate under specific conditions, allowing hackers to take control of self-driving cars with a 99% success rate while remaining invisible to current security measures.

Why it matters

The discovery of VillainNet highlights the serious security risks posed by AI-powered autonomous vehicles, which are becoming increasingly prevalent on public roads. If left unaddressed, this vulnerability could enable dangerous scenarios like hackers hijacking self-driving taxis and holding passengers hostage, threatening public safety.

The details

VillainNet exploits the architecture of modern AI systems used in self-driving cars, which rely on 'super networks' - massive systems that swap smaller modules in and out depending on the task. Researchers say an attacker only needs to poison one tiny tool in this digital toolbox, allowing malicious code to remain invisible across countless normal configurations until the car calls up that specific module, at which point the backdoor activates.

• The research was presented at a major security conference in October 2025.

What they’re saying

What’s next

The research team is calling for new defenses to be developed by automakers before these types of attacks move from the lab to public roads.