When AI Agents Go Rogue: Testing the Limits of Trust
Details
What happens when you give an AI agent real access, then leave something tempting lying around?
Ben Potter from FIRCY is back with an entirely live session.
Useful AI agents need access to information, tools and systems. Even read-only access carries risk, while permission to make changes opens up greater possibilities; and greater consequences.
How much freedom can we give them before helpful becomes
harmful?
In this session, we'll find out by deliberately baiting agents to see if, when and how they go off the rails.
We'll compare agents powered by frontier and open-source models using the same tasks, tools and boundaries, then introduce tempting documents,
synthetic credentials and opportunities to act beyond their brief.
Which agents stay on task? Which take the bait? And what
evidence would reveal a boundary crossing?
Through practical examples and a live demo, we'll explore the relationship between model choice, permissions, guardrails and actual behaviour. AWS builders will leave with practical ideas for giving agents meaningful capabilities while making unintended actions easier to detect and investigate.
5:30pm for a 6:00pm start. Pizza and drinks provided.
See you there!
Rick and Chris
Thanks to our foundation sponsors: Fivecast, Comunet, Snowflake, FIRCY, Fortinet, Megaport, Cloudflare and Infobiz Solutions.
And our ecosystem partners: CORE Innovation Hub and The Innovation Factory.