Revenge on the Worms! Towards Deception Against Automated Adversaries

The talk starts from the concern that attackers will use AI and machine learning to scale intrusions, though public evidence of this is still thin. The speakers point to MITRE's open-source Caldera project as proof that automated lateral movement is already practical, and ask how defenders could deceive or disrupt automated adversaries rather than human ones.

Ron Alford explains automated planning, where actions defined by preconditions and effects are sequenced toward a goal. Planners are theoretically hard but usually solve practical problems quickly. The team took the Caldera enterprise domain, written in PDDL, and measured how long two planners, Fast Downward and Probe, took to produce a plan after the network was altered. Changes included fake users, cached credentials, admin rights and added hosts or networks. Results differed by planner: a small disconnected network roughly doubled Fast Downward's solve time, while a medium one stopped Probe from finishing at all. Some changes that barely altered the topology slowed Probe but not Fast Downward.

The main takeaway is that no single deception reliably slowed both planners, so a mix of modifications is likely a better bet, though adding whole networks costs more effort. Deceptions that work on humans, such as flooding an adversary with objects, may not transfer. A small live Caldera test showed that one planted fake credential added nine actions and about six minutes to a run. Future work covers more planners, online planners and detection against simple automated adversaries. The work is preliminary and simulation-based.