A list of ten numbered categories is not yet mathematics. Nor are arrows between those categories. It gets interesting where terms acquire a fixed meaning, assignment rules become explicit, and a model yields statements that can be checked.
That is exactly where the Top Level Cyber Threat Clusters (TLCTC) meet applied mathematics.
TLCTC can serve as the semantic foundation for mathematical models of cyber attacks, and its classification structure can be described formally. Together with the companion paper Layered Event Modelling: Semantic Separation as a Precondition for Explicit Causal Analysis, the work already goes further: that paper contains a first explicit formalisation of layers, relations and causal transition durations.
The two papers do different jobs. TLCTC classifies attack steps by the generic vulnerability each one exploits. Layered Event Modelling structures events, states and relations. Connected, they let questions about attack paths, downstream consequences and points of intervention be stated more precisely.
This post explains the connection. The classification formulas and the worked example are presented for this post; where the existing formalisation is taken up or extended, that is said explicitly. None of it is a new normative TLCTC specification. The normative source remains the v2.5 framework dictionary and the core paper.
First, Fix What Is Being Classified
Anyone who wants to analyse cyber attacks needs a clear unit of analysis. A complete incident can involve many actions, several systems and different consequences. Giving it a single label collapses that structure.
TLCTC starts from the individual attack step, classified by the generic vulnerability it exploits. The core paper supplies the cluster definitions, boundary tests and precedence rules for this, and it keeps the classification of the cause apart from events and consequences. [core paper §§3, 4 and 6]
This raises a first mathematical question: can every sufficiently described, atomic attack step be assigned exactly one cluster?
Let be the set of attack steps within an explicitly declared scope. In TLCTC that scope is settled before any cluster is considered: only steps in the Attack row of the cause-side partition are classified, so failures, errors in use and abuse of rights never enter the set. [R-SCOPE, core §3.5] Let
be the set of cluster identifiers. The intended classification is a function
The statement behind it is simple: for every step, the classification yields exactly one cluster identifier. It is Axiom VI, the single-cluster rule, written as a function.
The word atomic carries a lot of weight here. A log entry, a process start or a single command is not automatically an atomic unit of the model. One technical event can contain several mechanisms that have to be told apart. Conversely, several observations can evidence the same modelled step.
Reproducible classification therefore also needs rules for decomposition. If two analysts split the same incident differently, their cluster counts can diverge even though both use the same cluster definitions. TLCTC's own rule is blunt: a step that appears to belong to two clusters is two steps, and must be split. [core §3.2]
Ten Clusters as a Partition: A Precise Claim
For each cluster, define the set of steps assigned to it:
A complete, non-overlapping division requires
and
The first condition says every step in scope is captured. The second says no step is assigned to two clusters at once. If all ten subsets are non-empty, they form a partition into ten classes.
One distinction is decisive. Once a total, single-valued function is assumed, these properties of its preimages follow mathematically. The hard work is showing that the domain rules actually yield such a function over the claimed scope. The formula does not do that work.
Nor does it follow that nature supplies exactly ten independently recognisable attack categories. Observable properties can overlap. A classification can still produce unique results if it resolves those overlaps with traceable rules.
The core paper states this distinction explicitly. Uniqueness is produced by the generic-vulnerability definitions, the boundary tests and the precedence rules: raw descriptions may genuinely overlap, and it is rules such as R-CHANNEL that make the assignment unique. Completeness is treated as a falsifiable hypothesis, not a theorem. [core §§3.2 and 8]
A robust test therefore has to look for counterexamples. Is there an in-scope attack step that cannot be assigned to any cluster even with sufficient information? Are there contradictory rules? Does the decomposition stay stable across independent analysts? The core paper issues that challenge in so many words: produce a step that cannot be decomposed into the existing clusters without redefining it.
Why #7 Three Times Can Be Three Different Steps
A multi-stage malware chain shows what the abstraction is for. Take a simplified example in which a stager executes, then a loader, then a further payload module:
For these three separately documented executions:
At cluster level this reads:
The repeated identifier is not redundant. It says that different steps exploit the same generic vulnerability. Time, executed code, affected process and capabilities gained can all differ. TLCTC asks for exactly this: a #7 step is recorded at each moment foreign executable content executes. [R-EXEC]
The mapping is therefore deliberately many-to-one. It compresses information. The concrete technical sequence cannot be recovered from the cluster sequence alone; the descriptions and evidence of the individual steps are still needed.
Equally, not every downloaded file justifies another execution step. R-EXEC ties #7 to execution, not to arrival. Event boundaries have to be kept, or a cause classification degrades into a count of technical artefacts.
From Attack Path to State Model
Classification tells us which generic vulnerability a step exploits. To analyse possible continuations we also need each step's preconditions and results.
One possible representation uses states as nodes and attack steps as labelled transitions:
A state might record which systems are reachable, which credentials are held and which execution capabilities exist. A step changes part of that state.
The benefit shows in a simple question: is the result of an earlier step enough to carry out the next one?
An acquired credential enables authentication only if further conditions hold. It must fit the target, still be valid and pass the required checks. TLCTC already separates the two moves: acquisition maps to the enabling cluster, and using a stolen credential to authenticate is #4. [Axiom X, R-CRED] But the arrow into #4 still says nothing about whether the preconditions for that use were met.
A modelled transition therefore needs at least three things described: its preconditions, the relevant action and its effect. Several preconditions can be jointly necessary. A simple picture with one incoming arrow must not hide that AND condition; a sufficiently complete state vector can hold it.
Such approaches have an established research base. The Iota project, for example, models the preconditions and effects of exploits and uses them to generate attack graphs for IoT systems. TLCTC could add a shared cause classification to models like that. The connection is a conclusion of this post, not an integration examined in the Iota paper. [Fang et al., Iota, IoTDI 2022]
What Nodes and Arrows Mean Belongs to the Model
A directed graph is, first of all, nodes and directed connections. Its mathematical structure does not fix what those elements mean in the domain. A connection might express an attack step, a classification or a piece of evidence. For analysis, that meaning has to be explicit.
The Layered Event Modelling paper separates the substrate hierarchy, the event-abstraction hierarchy, the representation hierarchy and the epistemic access structure; since draft v0.6 it adds a fifth, normative structure. It distinguishes causal, constitutive, classificatory, representational and evidential edges, and since v0.6 a normative edge type as well. Architectural dependencies are kept in a separate, type-level stratum. [Layered Event Modelling §§4–6]
The examples below illustrate the edge types for a payment environment:
| Edge type | Example | What an analysis method must respect |
|---|---|---|
| Causal | A malicious module changes the amount of a payment instruction. | A mechanism changes a state; its preconditions and possible interruptions matter. |
| Constitutive | The individual account balances at one point in time constitute the aggregated balance. | The aggregation describes what the overall state consists of; here it is not a later event. |
| Classificatory | A documented execution of foreign executable content is assigned to cluster #7. | The assignment applies criteria; it does not cause the execution. |
| Representational | A payment instruction represents an intended payment order. | What must be checked is whether representation and referent correspond. |
| Evidential | A process log supports the finding that a malicious module was executed. | Evidence and the evidenced event have to be kept apart. |
These examples add no TLCTC rules. They show why a path algorithm must not run indiscriminately over every arrow in a domain model. Otherwise a link from a malicious module to a log entry, and from there to a cluster identifier, could wrongly appear as attack progress.
State models and event models must likewise stay distinguishable. In the notation introduced above, states are nodes and attack steps are transitions. In an event graph, by contrast, the executions that occurred can themselves be nodes. Both representations are usable; their nodes and edges mean different things and need an explicit mapping between them.
An Existing Formalisation, with Explicit Extensions
Section 12 of the Layered Event Modelling paper proposes a typed weighted multigraph. Nodes are separated by layer, and edges carry types. The duration function is restricted to causal edges. That is already mathematical modelling, even though domain validation is still outstanding.
For this post the representation can be sharpened so that several edges of the same type between the same nodes stay distinguishable. Let be the node set, a set of individually identified edges and the set of edge types. We use:
Each edge thus gets a source node, a target node and a type. Two edges may share endpoints and type while denoting, for example, different documented mechanisms. These explicit edge identities are a refinement of the paper's compact triple notation.
For the node assignment, let be the set of admissible pairs of structure and layer. Then:
Every model node belongs to exactly one such pair, as the paper's axiom A1 requires. That prevents a single node from having to express a process, a data content and its business meaning at once. Where different nodes describe the same real-world occurrence from different perspectives, that co-reference is recorded as a separate relation.
The TLCTC assignment remains a separate function over attack steps. It is not applied indiscriminately to every edge: a piece of evidence, a data consequence or a class assignment is not an additional attack step.
This yields an important division of labour. The graph structure describes which elements are connected. The edge types determine what those connections mean. The cluster assignment names the generic vulnerability of each documented attack step.
Causes, Data Consequences and Time Must Stay Distinguishable
A cause classification must not change its meaning because of a later consequence. If a malicious module executes and then encrypts data, the mechanism and the data impairment must remain separately documentable.
In this example the execution is classified as #7. The loss of the data's accessibility is recorded as a Data Risk Event, [DRE: Ac]: the data is present but unusable. The same execution could have had other consequences, or none detected at first. This is Axiom III at work: threats are causes, not outcomes, and "ransomware" is never a cluster.
Such data states are represented separately in the companion event model. That avoids asking one cluster identifier to express mechanism, extent of damage and business impact at once. A consequence at business level needs its own description as well; in TLCTC terms it is a Business Risk Event.
One caution applies: assigning nodes to different layers does not by itself establish a causal transition. Every claimed consequence needs a stated mechanism. The reading direction of a diagram does not replace that check.
The calculation
gives, in the first place, the interval between two fixed points in time. It does not establish causation. The Layered Event Modelling paper therefore defines a causal transition duration only for causal edges: [Layered Event Modelling §§6 and 12]
TLCTC's own velocity annotation keeps the same discipline: →[Δt=…] sits on the sequence operator between two attack steps, never on a classification. [core §7.2]
For a classificatory relation such a duration is undefined, which is not the same as a duration of zero. If an analyst needs three hours to make an assignment, that describes a separate work process. It does not turn the relation between event and category into a three-hour causal transmission.
Nor may an unresolved transition simply be set to zero. Identical timestamps can result from limited measurement resolution, and the analysis should then record the temporal uncertainty. TLCTC notation already admits bounded values such as Δt=<10m.
For response options, at least three points in time have to be kept apart: when did the relevant event occur, when was it detected, and when did an effective intervention take effect? The time available also depends on which specific follow-on step is to be prevented. The core paper makes the same comparison for attack-path edges with the Detection Coverage Score, DCS = MTTD / Δt. [core §7.2]
Is There Time to Block the Payment?
The manipulation occurs at second 2 and is detected at second 5. The payment is scheduled for second 10. Change the time between detection and an effective block.
The block takes effect at second 7, 3 seconds before the scheduled payment. The DRE has already occurred; the manipulated payment order is not executed.
Illustrative model. Without a timely block, the manipulated payment is executed at second 10. If block and payment fall on the same second, the order is undetermined. That the block works is assumed; real controls and timings need their own evidence.
A Payment Example: Event Time Is Not Detection Time
The following example is entirely illustrative. It describes a modelled sequence, not measured company data.
| Time | Event or observation | Meaning in the example |
|---|---|---|
| 10:00:00 | Downloaded malicious code executes. | Attack step #7; a System Risk Event is recorded. |
| 10:00:02 | The code alters a payment instruction. | A Data Risk Event has occurred, [DRE: Ii]: the instruction no longer corresponds to the intended payment order. |
| 10:00:05 | Monitoring detects the manipulation. | The DRE that has already occurred is observed. |
| 10:00:07 | A block prevents execution of the affected payment order. | An intervention interrupts the assumed further sequence. |
| 10:00:10 | Execution of the payment order was scheduled for this time. | The prevented follow-on step is not an event that occurred. |
Detection at 10:00:05 cannot retroactively prevent the DRE at 10:00:02. It can, however, trigger a response that takes effect before the scheduled payment. In the example, the execution of the manipulated payment order, the Business Risk Event, therefore does not occur.
No observed interval between DRE and BRE may be claimed here, because the BRE is precisely what did not happen. What can be measured is the time to detection and the time to an effective block. The remaining time until the scheduled execution is a further, separate quantity. Whether the payment would actually have been executed without intervention remains a statement about the counterfactual course under the model's assumptions.
This is what the formal separation achieves in practice. It distinguishes a successful compromise, a data consequence that occurred, an observation and a prevented business impact. And it stops an averted loss from being read as evidence that no attack succeeded: the SRE and the DRE stand, whatever happened to the BRE.
Dependencies Determine Which Consequences Become Possible
For the payment example, the decisive question is under what conditions an altered instruction can reach execution at all. Is it processed immediately? Is there independent approval? Is the order reconciled against a trusted source before execution?
These conditions are part of the architecture and the process logic, and they existed before the specific attack. The Layered Event Modelling paper separates such dependencies from individual events that occur; their formal connection to the event graph remains an open task there. [Layered Event Modelling §6.1]
In an application, a transition rule might require that an executable instruction, an approval and an open execution window all hold at once. A control that changes these preconditions changes the possible continuations.
A successful control therefore need not always stop an event on an existing path. An architectural change can already make a particular follow-on step impossible under the previous conditions. The analysis model has to capture that change in preconditions.
Where This Becomes Applied Mathematics
Once states, transitions and constraints are sufficiently described, concrete mathematical questions become tractable. Each time, it has to be fixed which part of the overall model is being evaluated. A cause analysis uses different relations than a search for available evidence.
| Question | Possible mathematical analysis |
|---|---|
| Is a given target state reachable? | Reachability analysis under the defined transition rules |
| Which sequence needs the fewest steps? | Shortest paths under a defined step count |
| Which interventions cut every modelled path to the target? | Graph-cut or optimisation problems under a defined control effect |
| Which paths remain possible after an intervention? | Repeated reachability analysis |
| How do delays change the options for intervention? | Analysis of timed state models |
The restrictions belong to each statement. The shortest path is not automatically the most likely one. A connection cut in a model does not automatically mean that a real control is fully effective. Where several preconditions are jointly required, the chosen method must respect their AND combination; an ordinary path through a graph with individual precondition nodes is not automatically enough.
And a model can only account for the paths its assumptions and information represent. The statement "all paths cut" therefore always has to say which model it refers to.
Quantitative risk calculation needs further data. Ten clusters do not supply ten universal probabilities of occurrence. Anyone chaining probabilities along a path has to account for their dependencies; a product of unconditional single probabilities presupposes additional, often strong, independence assumptions.
This is the step into applied mathematics: a domain problem is translated into an explicit structure, examined with suitable methods and then interpreted under its assumptions.
Unknown Is Not an Eleventh Cluster
A complete classification model would still not guarantee complete observation of an incident. Missing logs, unclear execution paths or contradictory reports can prevent a sound assignment.
Mathematically these are different levels. The intended classification function concerns sufficiently determined attack steps. The actual analysis works with limited evidence and may therefore remain provisionally incomplete.
TLCTC's notation keeps the two levels apart. A step that evidently occurred but cannot yet be classified is written ?; a gap of unknown length is written …. Both are epistemic annotations, not clusters: they are excluded from statistics and never carry a Data Risk Event. [R-UNRES-2, R-UNRES-5, core §7.7] In the terms used above, a ? marks a step whose value under cannot yet be read from the evidence. The function is not in doubt; our knowledge of its argument is. And where any cluster can be defended, even weakly, the notation asks for #X [conf=low] instead. [R-UNRES-9]
Mixing these levels turns information gaps into apparent certainty. An automated system must be allowed to preserve uncertainty. A forced cluster assignment improves neither the data set nor the model.
This matters for AI applications too. Explicit definitions and rules can make machine outputs more checkable. They guarantee neither correct interpretation nor safe decisions. A model can emit formally valid identifiers and still have misunderstood the underlying facts.
What Is Already Formalised, and What Remains to Be Tested
Mathematical notation makes assumptions visible and consequences derivable. It cannot replace the correctness of the starting concepts or the completeness of the observations.
The existing formalisation is therefore concrete progress: it fixes which model components and relations are admissible. Empirical corroboration is a further task. The Layered Event Modelling paper names open completeness questions and a still-outstanding test of its diagnostics by independent raters. [Layered Event Modelling §§13 and 14]
For the connection with TLCTC this yields several separate test tasks. Do analysts agree on decomposition and cluster assignment? Do they distinguish the same edge types? Do predicted preconditions match observed sequences? And does the model improve the selection of effective interventions?
These questions must not be collapsed into a single metric. High agreement between analysts first shows reproducibility. It does not yet prove that a model captures all relevant mechanisms or describes the effect of an intervention correctly.
The completeness of the ten clusters also remains a question of its own. Formally clean event modelling can support that claim by making counterexamples easier to state. It does not prove it by merely existing.
A Name That Holds
So is TLCTC applied mathematics?
Its cause classification has a structure that can be described mathematically. Together with Layered Event Modelling, a first explicit foundation for formal models of events and relations already exists. Path, state and intervention analyses built on it are applications of mathematical methods to cyber problems.
For TLCTC on its own, cause-oriented classification model remains the precise name. For the combination with Layered Event Modelling one can already speak of an approach to the formal modelling of cyber and risk events. The step into mathematical modelling has begun; its practical power has to be examined on concrete cases.
One possible positioning:
TLCTC supplies the cause-oriented classification of cyber attack steps. Layered Event Modelling adds an explicit formal structure for events, states and relations. Together they form a foundation for mathematical analysis of attack paths, downstream consequences and points of intervention.
The value of such a foundation shows in the questions it makes testable. Was the same generic vulnerability classified consistently? Are the preconditions of a follow-on step met? Which paths remain open? Which statement survives a counterexample?
Mathematics can only work with the distinctions we actually draw in the model. Defining those distinctions cleanly is already a substantial part of the work.
Further Reading
- Layered Event Modelling: Semantic Separation as a Precondition for Explicit Causal Analysis — the structures, edge types and causal durations this post builds on
- Core paper — non-overlap (§3.2), the cause-side partition (§3.5), classification rules (§6), attack-path notation and Δt (§7), limitations (§8)
- Framework dictionary (JSON) — the normative cluster definitions, axioms and rules
- Attack Velocity — Δt, the velocity classes and the Detection Coverage Score
- Fang et al., “Iota: A Framework for Analyzing System-Level Security of IoTs”, IoTDI 2022