Threat actors can poison AI models by tampering with their underlying infrastructure, according to a report from security firm Eclypsium.
The vulnerability or weakness
Compromising firmware and software powering the computers running AI models provides a path for adversaries to change model weights.
Earlier attacks and coverage
No earlier related attack in our archive yet.
Business impact
This could lead to biased model outcomes, weakening certificate validation or selecting unsafe dependencies, impacting operations and potentially causing legal issues.
Industries affected
AI labs, tech companies, and organizations relying on AI models are at risk.
Users affected
Users may face inaccurate or biased model outputs, affecting decision-making processes and customer trust.
Build these skills
Watch the daily brief
Sources
- Hackers could poison AI models by tampering with core infrastructure (Cybersecurity Dive)
