Not every rogue agent is an attack
Tuesday, November 17, 202611:00 AM ET
An agent doesn’t need an attacker to do something you wouldn’t have approved.
Join Origin researcher Tyler Holmwood for a live walkthrough of how agents go off course. Follow real traces to see what the agent was asked to do, which instructions it followed, and how its actions crossed a line.
See how to recognize risky behavior, review what your agents can access, and use controls you already own to limit the consequences.
Tyler Holmwood
Staff Security Researcher
Origin Technology
Tyler researches how coding agents behave on real endpoints. His most recent work reviews their automatic approval modes, the settings that let an agent act without asking first.
What we’ll examine
From a prompt to actions on your machine
How an agent moves from a request to tool calls, commands, and changes on an endpoint.
The access and authority behind those actions
What an agent can reach, whose credentials it uses, and where its instructions come from.
Two paths to risky behavior
Outside instructions that redirect the work, including prompt injection and tampered skills. And an agent’s own decisions that take a legitimate task beyond what you intended.
What the traces reveal, and what you can do
Follow what agents saw, did, and changed. Learn where to look for risk and which existing controls can limit the consequences.
By Tyler Holmwood