Discover

● AI and generative AI systems can be easily hijacked to generate malicious code, even when designed to reject such requests.
● Other types of attacks, known as "model evasion attacks," exploit modified inputs to cause unexpected behaviours in AIs, such as making a self-driving car misinterpret traffic signs.
● Poisoned data can introduce backdoors into AI models, leading to unintended behaviours, which is concerning due to the lack of control engineers have over their data sources.