Speaker: Muneeb Zafar 

AI agents are moving from systems that generate advice to systems that can act through tools, APIs and connected environments. In offensive security, that shift is especially visible: an agent may be capable of reconnaissance, validation and multi-step attack reasoning, but capability alone does not tell us what it should be allowed to do.

This session introduces the idea of an “autonomy envelope”: externally enforced boundaries around authority, execution environment, side effects, oversight and stop conditions. Using offensive security as a stress test, Muneeb Zafar will show why model-level safety is not enough once AI can act, and why trustworthy autonomy depends on the system architecture around the model.

The principle is simple: maxim ...