The One About Jailbreaks & Kill Switches

How do you find when AI systems cross their intended boundaries, and stop agents from doing things they shouldn't? Join practitioners from GovTech Singapore and Salesforce as they share two different approaches to keeping AI systems in check, from using red-teaming agents to uncover policy-breaking chatbot behaviour, to controlling how agents access and act across tools and systems. More About The Sharings Watson Chua (Staff Data Scientist, GovTech Singapore) will share on “Breaking Chatbots with Red-Teaming Agents.” How do you find weaknesses in an AI chatbot that conventional evaluations might miss? Drawing on GovTech's work red-teaming chat.gov.sg, Watson will share how his team built AI agents to simulate adversarial conversations and uncover behaviours that violate the chatbot's intended policies. Explore how the team orchestrated multi-turn attacks, experimented with different jailbreaking techniques, and evaluated successful attempts, including the challenge of distinguishing genuine policy violations from false positives. Hear what they learnt from using agents to probe AI systems and how those findings informed stronger safeguards. (Technical Level: 200) Aksel Yap (Public Sector Solution Engineering Lead for Data Foundations, Salesforce) will share on “One Place to Govern Them All: Safe Access, Right Outcomes and Governance for Your AI.” AI agents and the models behind them change quickly, but the systems they need to interact with often don't. As organisations deploy more agents across different platforms, how do you give them access to existing systems without creating a separate set of connections, permissions and controls for each one? Using MuleSoft Agent Fabric as a working example, Aksel will demonstrate how existing APIs can be exposed as scoped tools for agents, and how a shared gateway layer can help teams discover agents, apply common policies, quarantine a runaway agent, and remove destructive actions from an MCP tool. He'll also explore how th