<ul><li>Nation-states and sophisticated threat actors can compromise cybersecurity by poisoning widely used open-source AI models, creating blind spots in critical infrastructure and government agencies.</li><li>Two categories of attacks on AI models are pre-training attacks involving embedding malicious data during training and post-training attacks executed after the model is deployed.</li><li>Examples of pre-training attacks include Model Poisoning, where tainted data trains the AI to ignore certain attack signatures, and Adversarial Evasion, where crafted inputs exploit model vulnerabilities.</li><li>Post-training attacks involve reverse-engineering detection capabilities and exploiting predictable blind spots in the model, highlighting the importance of rigorous adversarial testing before integrating open-source AI into security solutions.</li></ul>

Open-source AI models are revolutionising cybersecurity, powering everything from scam/fraud…

Discover more