The One About Jailbreaks & Kill Switches

GovTech Singapore and Salesforce staff give two talks at Lorong AI @ One-North on Wed 11 Nov 2026, both circling the same worry: what an AI system does when it goes off-script. The first is about red-teaming agents — AI built to hold adversarial conversations with a chatbot and draw out behaviour its own policies forbid. The work draws directly on GovTech's red-teaming of chat.gov.sg.

It gets into the mechanics rather than staying on principle: running multi-turn attacks, trying out different jailbreaking methods, and finding flaws in a chatbot that standard evaluations let through. Then the awkward part — assessing which attacks actually landed, and how hard it is to separate a genuine policy breach from a false positive. The team also covers what using agents to test AI systems taught them, and how those findings fed into stronger safeguards.

The second talk switches from testing to access: how organisations let AI agents into their existing systems without setting up separate connections, permissions and controls for every platform. Both talks run back to back from 3pm to 5pm, and entry is free.