Securing AI and agents

AI systems introduce new ways to fail: manipulated prompts, poisoned data, leaked models and, increasingly, software agents that take actions on their own. This hub looks at how to secure all of it. The focus is on what a security team can do in practice: what to test before deployment, how to limit an agent's permissions, and which questions to put to vendors. The aim is to use AI without handing attackers a new way in. Topics include digital agents, jailbreaks and prompt injection.