A further risk comes from what’s known as adversarial AI, where attackers deliberately try to trick or disrupt AI systems. This is no longer just a theoretical issue – cyber criminals are already using techniques like prompt injection, data manipulation, and exploiting weaknesses in AI models to bypass security measures or undermine the reliability of AI-powered defences.
As AI becomes more widely adopted, these kinds of attacks are expected to become a growing concern for organisations.
Another threat is that AI could be used to undermine itself.
Adversarial AI (not to be confused with generative adversarial networks, or GANs, which are completely different) involves confusing and misleading AI systems under real-world conditions. A variety of techniques are already possible, including:
- Tricking machine learning models by feeding subtly deceptive data so that they misclassify images or spam.
- Data poisoning attacks involve manipulating the data fed to the AI model to degrade its performance or introduce vulnerabilities (this requires access to the data).
- Fooling online classification systems by sending them fake data.
- A denial of service attack where an AI is slowed by being bombarded with complex problems (this is a newer and less common threat compared to other adversarial techniques).
Another factor increasing the vulnerability of AI to adversarial attacks is that many won’t be centralised. Edge computing is a major trend in network design. It’s based on the idea that servers and applications should be placed closer to the users or devices they serve, helping to improve performance and make systems more resilient.
This also applies to AI located at edge locations, which could make them harder to monitor for tampering and manipulation.
