An arXiv preprint posted 24 July 2026 by a University of Adelaide doctoral candidate proposes deploying decoys and admin-feedback loops to harden Active Directory, moving beyond the patch-everything default that dominates AD defence S¹. The thesis argues existing security models assume attack graphs stay frozen when real AD environments shift by the hour. What the title calls "AI-Driven" turns out to be optimisation and game theory, and the authors concede the underlying maths is computationally intractable. Whether the ideas survive contact with a real enterprise network is the question the preprint does not answer.

My read: The honest admission that these problems are intractable is the most interesting line in the whole abstract. Most security papers bury their limitations; this one puts computational intractability front and centre. I don't buy the "AI-Driven" framing, because the methods described are graph optimisation and game-theoretic decoy placement, not anything I'd call AI in 2026. The admin-feedback loop idea, though, is the kind of practical detail I haven't seen in AD security research: acknowledging that not every fix can be implemented, and building prioritisation around that constraint. Whether the optimisation approximations they must be using to handle intractability hold up at enterprise scale is what I'd watch for.

Why static graphs miss the attack

Active Directory is the system that manages who can access what across a Windows domain network S¹. It is also, as the authors note, a primary target for adversaries, because compromising AD means controlling the entire domain S¹. Security researchers model AD as an attack graph: nodes are accounts and computers, edges are the permissions and trust relationships an attacker can exploit to move laterally.

The thesis identifies a core flaw in how these graphs get used for defence. Existing models treat the graph as a fixed snapshot S¹. Real AD environments are nothing like that. Users join and leave. Permissions get granted for a project and revoked when it ends, and trust relationships between domains shift as the organisation restructures. A path that exists on Monday may be gone by Wednesday. A defence plan built on a static graph is defending against a network that no longer exists.

Decoys instead of patches

Most proposed AD defence solutions focus on one move: removing edges, which means revoking the permissions that create attack paths S¹. The authors call this passive and argue that more active defence mechanisms are "largely unstudied" S¹.

Their alternative is a honeypot and decoy placement model. The idea is to position fake nodes in the graph so that an attacker traversing the network hits a decoy before reaching a Domain Admin account S¹. The placement strategy tries to shrink both the count of shortest routes an attacker could follow and the set of starting points from which a Domain Admin account is within reach S¹.

A second defence layer accounts for the dynamic nature of the graph. Rather than placing decoys once, the model looks for decoy positions that give defenders the longest possible reaction window, even as the graph keeps evolving S¹. In plain terms: put the decoy where it will still be in the attacker's way even after the network shifts.

The admin feedback problem

The thesis introduces an adaptive prioritisation model that does something unusual for a security paper: it asks the IT administrator about each high-risk attack path before acting S¹. The reasoning is practical. Not every remediation can be implemented. Some permissions exist for a business reason. Some edge removals would break a critical workflow.

The model then tries to minimise the admin's approval burden by finding a general policy that extends the admin's decisions on a few edges to other edges with similar risk features S¹. An admin says "yes, revoke this" three times, and the model learns to apply the same judgement to comparable paths without asking again.

The maths fights back

Here is where the thesis gets honest in a way most security papers do not. The authors state that every model they contribute rests on problems that are computationally intractable S¹. Decoy placement, dynamic defence, adaptive prioritisation: each one, in its exact form, cannot be solved efficiently as the graph grows.

This is not a footnote. An enterprise AD environment can have hundreds of thousands of nodes and millions of edges. If the exact optimisation is intractable, the model must rely on approximations or heuristics. The abstract does not describe what those approximations are, how close they come to optimal, or how they perform at scale.

What to do about it

For IT teams running Active Directory, the thesis points at a gap in how most organisations think about AD defence. The default is reactive: find a vulnerability, patch it, repeat. The ideas in this preprint suggest a different posture.

Consider a mid-sized law firm with 800 staff, 30 application servers, and an AD forest that has accumulated trust relationships over a decade of mergers and IT staff turnover. The security team runs a tool like BloodHound, finds 200 attack paths to Domain Admin, and starts revoking permissions. Half the revocations break something. The other half get re-granted a month later because someone needed access for a client matter. The cycle repeats.

The thesis describes exactly this failure mode and proposes a way out: prioritise based on what the admin can actually implement, place decoys to catch attackers who find paths you missed, and account for the fact that the graph will change before you finish fixing it. None of these ideas are available as products today. But the framing is useful for security teams evaluating tools. Does your AD defence tool assume the graph is static? Does it account for admin constraints? Does it offer anything beyond edge removal?

One practical thing to check this week: run BloodHound or a similar AD attack-path mapper on your domain twice, two weeks apart, and diff the results. If the attack paths have shifted, your static-graph defence model has a blind spot.

What we don't know yet

The preprint is a doctoral thesis from the University of Adelaide's School of Computer and Mathematical Sciences, supervised by Dr Hung Nguyen, not a peer-reviewed journal paper [S1, P2]. The abstract describes no empirical results, no real-world testing, and no deployment in a live enterprise environment S¹. The "AI-Driven" label in the title does not match the methods described in the abstract, which are optimisation-based and game-theoretic rather than neural or generative S¹. The arXiv categorisation includes q-fin.GN, a quantitative finance category that is unusual for a cybersecurity paper and unexplained S¹.

The computational intractability admission means the practical value of these models depends entirely on the quality of the approximations used, which the abstract does not detail.

The next signal: the full thesis text, which should contain the approximation methods and any experimental results, is available via the arXiv listing S¹. We'll check whether the empirical chapters validate the models at enterprise scale or whether the intractability proves fatal in practice. If you want that follow-up, subscribe and we'll send it your way.

Sources: S1 — Practical Graph Optimisation and AI-Driven Models for Active Directory · P2 — Contents · P3 — OpenADUC/openaduc · P4 — Zhubaoheng/Graph-GRPO · P5 — GitHub - Tencent-Hunyuan/Hy3: Hy3 (295B A21B), a leading reasoning and

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.