Decision Making in Adversarial Automation

Introduction: Stochastic Parrots vs. Deterministic Predators

Bobby Kuzma and Michael Odell ("Moch") delve into the mathematical and algorithmic challenges of automated decision-making in adversarial cybersecurity environments. The speakers examine the contrasting paradigms between "stochastic parrots" (generative AI models that generate probabilistic text and decisions based on statistical correlations) and "deterministic predators" (systems designed to execute reproducible, deterministic attack strategies based on formal state verification).

The Mathematical Challenge of State and Action Spaces

Automating adversarial behavior in enterprise networks introduces extreme computational complexity:

State Space Explosion: Enterprise environments consist of tens of thousands of hosts, user accounts, services, permissions, and network routes. Modeling every potential system state creates a combinatorial explosion that exceeds manageable compute capacity.

Action Space Complexity: At any given moment, an automated agent can choose from thousands of potential tools, commands, and exploit parameters. Evaluating every state-action pair (Q-values) becomes intractable under traditional reinforcement learning without rigorous constraints.

Limitations of Rule-Based State Machines in Adversarial Testing

While traditional checklists and decision trees are algorithmically straightforward and explainable, they function as rigid state machines. In dynamic network environments, static decision trees fail to adapt when unexpected conditions arise, and their deterministic nature makes them easily predictable and signature-detectable by modern defense mechanisms.

Graph-Based Modeling and Subgraph Matching

To navigate massive network topologies efficiently, the speakers borrow graph theory principles popularized by BloodHound and financial fraud detection systems:

Property Graphs: Representing the target network, systems, credentials, and access relationships as rich property graphs.

Subgraph Matching: Instead of searching the entire graph at every step, automated engines evaluate whether local subgraphs meet the exact prerequisites required to execute a specific offensive technique.

Patterned Preconditions: Techniques and tools are modeled with explicit input requirements (e.g. valid SMB credentials, port 445 reachability, local admin rights) and expected outcome states.

Tooling Architecture: Property Graphs, DSL, and Kùzu

Kuzma and Odell introduce an open-source tooling architecture developed for deterministic TTP selection:

Domain Specific Language (DSL): A YAML-based DSL defining preconditions, graph query patterns, tool arguments, and graph transformations upon execution.

Embedded Graph Engine (Kùzu): Rather than deploying resource-heavy client-server graph databases like Neo4j, the architecture utilizes Kùzu—an open-source, embedded, in-process property graph database built for fast Cypher queries and analytical workloads.

Demonstration: Telemetry-Driven Attack Execution

The presenters demonstrate a live-recorded simulation where an operator executes basic reconnaissance commands (such as nmap and Linux enumeration). The engine ingests operator console telemetry in real time, continuously updates the local Kùzu property graph, evaluates matching subgraphs, and autonomously suggests the next optimal attack vector—guiding the workflow through hash extraction, hash cracking, and Active Directory domain dumping (secretsdump).

Key Takeaways and Future Horizons

Structure Before Automation: Effective adversarial automation requires robust, graph-based structural representations of target environments before applying AI or probabilistic logic.

Deterministic Explainability: Deterministic graph engines enable operators to explain exactly why a particular action was taken at any given second.

Open-Source Tooling: The speakers encourage the offensive and defensive communities to leverage their open-source DSL and embedded graph framework to build scalable, explainable security automation.