INTERFACE Seattle
Presented by Greg Anderson, Head of Solutions Engineering • Sprocket Security
Every autonomous pentest vendor now ships agents that run reconnaissance, exploitation, and validation against live environments. Almost none can tell a buyer what those agents are forbidden to do. The AI governance most organizations already hold does not help, because it rests on an assumption that all harm should be avoided. For a penetration testing agent, controlled, authorized harm is the product. A framework built to stop all harm cannot say how to authorize the right harm, in the right proportion, against the right target.
This session reframes the safety question for the people who buy autonomous offensive security rather than build it. It separates controls that merely instruct an agent to behave from controls that make misbehavior structurally impossible, the difference between “probably will not” and “cannot”. Attendees leave with a scorecard they can run against any vendor, including the one presenting, scored on evidence rather than assurances.