AI agents face growing prompt injection risks
Experts say attackers increasingly use social engineering to manipulate AI systems.
AI developers are working on new defences against prompt-injection attacks that aim to manipulate AI agents. Security specialists warn that attackers are increasingly using social engineering techniques to influence AI systems that interact with online content.
Researchers say AI agents that browse the web or handle user tasks face growing risks from hidden instructions embedded in emails or websites. Experts in the US note that attackers often attempt to trick AI into revealing sensitive information.
Engineers are responding by designing systems that limit the impact of manipulation attempts. Developers in the US say AI tools must include safeguards preventing sensitive data from being transmitted without user approval.
Security teams are also introducing technologies that detect risky actions and prompt users for confirmation. Specialists argue that strong system design and user oversight will remain essential as AI agents gain more autonomy.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!