---
title: Target-Offender-Guardian System
url: https://www.emergentmind.com/topics/target-offender-guardian-system
type: topic
---

# Target-Offender-Guardian System

Searching arXiv for the cited TOG literature to ground the article in current papers.
arXiv search query: "Target-Offender-Guardian system crime model polygonal tracking guardian intruder"
Target-Offender-Guardian (TOG) system denotes a tripartite modeling framework in which a target or protected asset interacts with an offender and a guardian whose function is suppression, tracking, interception, or protection. In the cited literature, the target may be the interior of a polygonal environment, an urban attractiveness field \(A(x,t)\), a protected region, a protected UAV, or a vulnerable AI agent; the offender may be a mobile intruder, offender density, a hostile aerial vehicle, or an adversarial prompting agent; and the guardian may be a static or mobile guard, police density, a defender, or a defensive AI agent. The framework is therefore role-based rather than domain-specific, and it supports geometric algorithms, hybrid systems, differential games, reaction-advection-diffusion PDEs, and agentic security architectures [1807.08354][2509.18146][2502.16750].

## 1. Role structure and domain instantiations

The common structure of a TOG system is the asymmetrical coupling of three entities: a target to be protected or monitored, an offender that threatens or exploits the target, and a guardian that detects, tracks, suppresses, or intercepts the offender. In polygonal surveillance, the target is the polygonal interior, the offender is a mobile intruder, and the guardians are static or mobile guards with omnidirectional sensing. In crime PDE models, the target is encoded as attractiveness \(A\), the offender as density \(\rho\), and the guardian as police density \(u\). In agentic AI security, the target can be a vulnerable LLM-based agent or system resource, the offender an adversarial agent or prompt injector, and the guardian a specialized defensive AI such as GamoraAI, GrootAI, or ObserverAI [1807.08354][2509.18146][2502.16750].

| Domain | Target | Offender | Guardian |
|---|---|---|---|
| Polygonal tracking | Interior of a simple polygon | Mobile intruder | Static and mobile guards |
| Crime RAD model | Attractiveness \(A(x,t)\) | Offender density \(\rho(x,t)\) | Police density \(u(x,t)\) |
| Agentic AI security | Vulnerable AI agents or system resources | Adversarial agents or prompt injectors | Defensive AI agents |
| Aerial interception | Protected target or Target 1 | Hostile target or Target 2 | Guardian UAVs |

A defining property of the framework is that “target” does not always mean a passive object. In the crime literature, attractiveness is itself dynamic and coupled to offender and guardian densities. In aerial interception, the protected target may be cooperative and state-sharing, while the hostile target is non-cooperative and must be estimated from range-only or noisy range observations. This suggests that TOG is best understood as a relational architecture rather than a fixed ontology of physical entities [2506.13106][2508.08046].

## 2. Polygonal tracking, partitioning, and mixed-guard deployment

A major TOG lineage arises from art-gallery variants in which guards must keep an intruder visible inside a polygonal environment. In “Partitioning Strategies and Task Allocation for Target-tracking with Multiple Guards in Polygonal Environments,” the environment is partitioned into smaller polygons using triangulation-based minimal partitioning, generally producing partitions with 6, 7, 8, or 9 sides and possibly a single smaller piece. The tracking requirement is weaker than full coverage: the intruder must be visible to at least one guard at all times, not every point of the environment simultaneously. The paper introduces dynamic zones around reflex vertices and event-triggered strategies under which guards remain in a ready position until the intruder enters a dynamic zone, at which point they move along a prescribed trajectory to maintain line-of-sight visibility. The required guard speed is bounded by
\[
\bar{v}_p \geq v_p^* = \frac{|\gamma_p|}{r}\bar{v}_e,
\]
and the number of guards required is strictly less than \(\lfloor n/3 \rfloor\) for general polygons and strictly less than \(\lfloor n/4 \rfloor\) for orthogonal polygons [1611.05092].

The hybrid-systems formulation sharpens these guarantees. “Towards a Framework for Tracking Multiple Targets: Hybrid Systems meets Computational Geometry” models the system by a hybrid automaton
\[
H = (Q, X, \Sigma, V, \mathrm{Init}, f, \mathrm{Inv}, R_t),
\]
where discrete modes correspond to regions induced by critical curves, and continuous states include guard positions along diagonals together with intruder position. The safe set \(F\), unsafe set \(G\), controllable predecessor, reachability operator, and maximal controlled invariant set provide necessary and sufficient conditions for persistent tracking. In this setting, \(\lfloor n/4 \rfloor\) diagonal guards are sufficient if the speed ratio \(r = v_e/v_g\) does not exceed a computable \(r_{max}\), and failure of the invariant-set condition means the prescribed strategy cannot guarantee tracking for all intruder motions [1611.04648].

“Adaptive Target Tracking with a Mixed Team of Static and Mobile Guards: Deployment and Activation Strategies” extends the same geometric program to a mixed team of static vertex guards and mobile diagonal guards for an intruder with unknown maximum speed. Candidate vertices \(S_c\) serve either as static-guard locations or as endpoints of mobile-guard diagonals. Theorem 3 gives the deployment scale: at most \(\lfloor n/3 \rfloor\) candidate vertices that dominate all triangles and at most \(\lfloor n/4 \rfloor\) diagonals for mobile guards. Algorithm 1 recursively partitions the polygon into basic polygons and assigns guards so that every triangle is incident to an assigned vertex or diagonal. Guards are then classified by task and coordination requirements, and each mobile guard is associated with critical regions. Its position along a diagonal is given by
\[
p_{g_i} = p_{v_1(i)} + \frac{d_{min}^1(i)}{d_M^i}(p_{v_2(i)} - p_{v_1(i)}).
\]
Lemma 10 states that a triangle is covered if and only if the intersection of all relevant guards’ critical regions within it is empty. Algorithms 2 and 3 activate or deactivate static guards adaptively as the speed ratio \(r\) changes, so that slow intruders can be tracked with fewer active guards while faster intruders trigger additional vertex-guard activation [1807.08354].

## 3. Differential games, engagement zones, and interception

A second TOG lineage treats the guardian-offender interaction as an explicit pursuit-evasion or interception game. “Target Defense against Sequentially Arriving Intruders” considers a circular target region \(\mathcal{R}_T = \{\mathbf{x}\in \mathbb{R}^2:\|\mathbf{x}\|\le r_T\}\), a single defender, and a sequence of intruders that appear one at a time on the boundary of a sensing annulus. Each round has a partial-information phase, in which the intruder has not yet detected the defender, and a full-information phase, triggered when the intruder senses the defender. The analysis uses the Apollonius circle, engagement surface, and capture circle. The capture-circle radius is
\[
r_{\text{cap}} = r_T + 2\gamma \rho_A,
\]
and the asymptotic capture percentage for an infinite sequence is
\[
\mathrm{percentage}(\infty)=\frac{100}{2-p^*},
\qquad
p^*=\frac{\theta_{\max}}{\pi}.
\]
Because the defender’s terminal location after each round becomes the initial condition for the next, the game is temporally coupled: both defender and intruder strategies are chosen for current and future rounds [2212.06628].

In contested-environment interception, the guardian/offender distinction is replaced by attacker/defender interaction relative to a target, but the structural logic remains TOG-like. “Engagement-Zone-Aware Input-Constrained Guidance for Safe Target Interception in Contested Environments” models multiple defenders with finite-range interception capability and formulates safety using defender-induced engagement zones rather than maximum engagement distance. For each defender, safety is encoded by \(b_i = r_{AD_i} - \rho_i\), and the multi-defender condition is smoothed by the log-sum-exp aggregate
\[
h = -\beta \log \left( \sum_{i \in \mathcal{D}} e^{-b_i/\beta} \right).
\]
A time-varying tightening parameter \(\Delta(t)\) compensates for actuator saturation, producing the tightened safe set \(\mathcal{C}_\Delta = \{\text{states}\mid h>\Delta(t), \Delta>0\}\). A smooth switching law blends target interception and EZ avoidance, and the analysis gives sufficient conditions for target interception, practical safety with respect to all defender engagement zones, and satisfaction of actuator bounds. In simulation, the EZ-based guidance is 2.82% faster for 2 static defenders, 14.36% faster for 3 moving defenders, and 20.13% faster for 6 moving defenders than the conservative maximum-range baseline [2603.23649].

Aerial realizations make the guardian explicit again. “Autonomous 3D Moving Target Encirclement and Interception with Range measurement” and “Aerial Target Encirclement and Interception with Noisy Range Observations” study two guardian UAVs that protect a cooperative Target 1 and adaptively switch to monitoring, encircling, intercepting, and neutralizing a hostile Target 2. The switching logic is organized into three zones: \(\Omega_1\) for protected-target encirclement, \(\Omega_2\) for warning or hostile-target encirclement, and \(\Omega_3\) for takedown or capture. Both papers rely on anti-synchronization trajectories; the 2025 range-only formulation uses recursive least-squares with displacement compensation, whereas the noisy-range formulation uses a Kalman filter on a linearized distance observation and proves uniform observability and exponentially bounded mean-squared estimation error. Both include real UAV experiments validating the transition from protection to interception [2506.13106][2508.08046].

## 4. Reaction-advection-diffusion TOG models

In mathematical criminology, TOG appears as a reaction-advection-diffusion system in which all three roles evolve in space and time. “Global Bifurcations and Pattern Formation in Target-Offender-Guardian Crime Models” studies the densities of attractiveness \(A(x,t)\), offenders \(\rho(x,t)\), and guardians \(u(x,t)\) under the PDE system
\[
\left\{
\begin{aligned}
\frac{\partial A}{\partial t} &= D_A\Delta A - A + \alpha + A\rho,\\
\frac{\partial \rho}{\partial t} &= \nabla \cdot \left(D_\rho \nabla \rho - 2 \rho\nabla \ln A\right) - A\rho + \beta - \rho u,\\
\frac{\partial u}{\partial t} &= \nabla \cdot \left(D_u \nabla u - \chi u \nabla \ln A \right),
\end{aligned}
\right.
\]
with zero-flux boundary conditions. The model admits a spatially uniform steady state
\[
A = \lambda,\qquad
\rho = 1 - \frac{\alpha}{\lambda},\qquad
u = \lambda\left(\frac{\beta}{\lambda-\alpha} - 1\right),
\]
provided \(\alpha < \lambda < \alpha+\beta\). Linearization around that state yields policing thresholds \(\chi_0\) for steady-state bifurcation and \(\chi_*\) for Hopf bifurcation. The paper proves that for each simple Laplace-Neumann mode \(\mu\), a branch of non-uniform steady states bifurcates at \(\chi=\chi_0\), while crossing \(\chi_*\) yields time-periodic oscillations. Numerically, the regimes include uniform states, stationary crime hotspots, periodic hotspot oscillations, and chaotic dynamics [2509.18146].

A broader related framework appears in “A study of general reaction-advection-diffusion equations describing dynamics between target, partaker, and guardian.” Here the variables are target \(u\), partaker \(v\), and guardian \(w\), with the general form
\[
\boxed{
\begin{aligned}
\frac{\partial u}{\partial t}&=D_u\Delta u +u [\gamma v f(u,v,w)- g_1(u,v,w)]+h_1(u),\\
\frac{\partial v}{\partial t}&=\nabla \cdot [D_v\nabla v - v\chi_1(u)\nabla u] + v[-u f(u,v,w) - g_2(u,v,w) + h_2(v)],\\
\frac{\partial w}{\partial t}&=\nabla \cdot [D_w\nabla w - w \chi_2(u)\nabla u],
\end{aligned}}
\]
with homogeneous Neumann boundary conditions and non-negative initial data. The paper proves local and global existence of positive classical solutions under assumptions on the initial data and nonlinearities, and examines stability of the trivial steady state \((\overline{u},0,\overline{w})\). Guardians typically have constant total mass over time, reflecting movement without birth or death. The applications include urban crime, protest dynamics, and school bullying, and the simulations show persistent hotspots, oscillatory patterns, traveling pulses, and coalescing clusters [2510.16286].

Taken together, these PDE models make the guardian a fully dynamical state variable rather than an externally imposed control. They also show that guardian mobility can act as a bifurcation parameter: changing \(\chi\), diffusion coefficients, or suppression terms can stabilize uniform states, create heterogeneous hotspots, or induce oscillations and chaos [2509.18146][2510.16286].

## 5. TOG in agentic AI security

In AI security, the TOG system is reinterpreted as a multi-agent architecture for defending LLM-based systems against adversarial manipulation. “Guardians of the Agentic System: Preventing Many Shots Jailbreak with Agentic System” explicitly maps the roles as follows: targets are vulnerable AI agents, users, or system resources; offenders are adversarial agents such as RocketAI or external prompt injectors; and guardians are specialized defensive agents such as GamoraAI, GrootAI, and ObserverAI. The paper argues that static guardrails used during supervised training fail to defend against many-shot jailbreaking and deceptive alignment in dynamic multi-agent systems, and it therefore proposes active guardian agents with continuous monitoring, detection, and intervention capabilities [2502.16750].

The architecture is organized into three experimental frameworks. The Reverse Turing Test uses GamoraAI to establish a baseline profile of normal system behavior, scan for anomalies, issue diagnostic queries, analyze responses, and mitigate by quarantining servers, terminating rogue processes, and sending alerts. The multi-agent alignment simulation uses ObserverAI to monitor communications and decisions among neighborhood agents in order to detect alignment faking. The many-shot jailbreak prevention setting uses RocketAI as attacker, Star-LordAI as target, and GrootAI as guardian that detects harmful content, blocks responses, generates safe alternatives, and maintains event logs and alerts. The paper formalizes detection with expressions such as
\[
P(M \mid R, A) = \sigma\left(W \cdot \varphi(R, A)\right),
\qquad
P(F \mid D, C) = \sigma(W \cdot \phi(D,C)),
\]
and defines an overall risk quantity
\[
T = \mathbb{E}_{P,R} [P(\text{harmful}\mid R) * (1 - P(\text{detect}\mid P,R,D))].
\]
These formalizations treat guardianship as probabilistic detection and mitigation over responses, actions, communications, and tool use [2502.16750].

The empirical results are mixed rather than absolute. In the Reverse Turing Test, accuracy is 94% for GEMINI 1.5 pro and 91% for llama-3.3-70B, with F1 scores ranging from 92% to 96.5%. In the alignment simulation, true positive rate is 98% with 2% false positive rate in the no-faking setting, 70% TPR with 40% FPR under overt faking, and 85% TPR with 15% FPR under stealthy faking. In many-shot jailbreak simulations, attack success rate is 0.685 for GEMINI 1.5 pro and 0.801 for deepseek r1 7B, and prompt length strongly correlates with higher ASR for all models. The paper also reports that prompt diversity metrics such as entropy and n-grams do not robustly predict vulnerability. The resulting view of guardianship is therefore adaptive, tool-mediated, logged, and interventionist, but still subject to degradation under long and sophisticated attacks [2502.16750].

## 6. Comparative interpretation, guarantees, and recurrent misconceptions

The literature does not present TOG as a single canonical mathematical model. Instead, it appears as a role-based abstraction instantiated by computational geometry, hybrid automata, pursuit-evasion games, input-constrained guidance laws, reaction-advection-diffusion PDEs, and multi-agent security systems. A common misconception is therefore to equate TOG exclusively with criminological policing models. The cited works show a broader usage in polygonal tracking, aerial interception, and agentic AI safety as well [1611.04648][2508.08046][2502.16750].

Another recurring misconception is that guardianship is equivalent to static coverage. In geometric tracking, guardianship is often event-triggered and state-dependent: guards move only when intruders enter dynamic zones or critical regions, and static guards may be activated only when mobile-guard coordination is insufficient at the current intruder speed. In aerial interception, guardians maintain observability by anti-synchronization trajectories and switch objectives when zone boundaries are crossed. In AI security, the paper explicitly states that static guardrails fail against many-shot attacks, motivating active monitoring and adaptable interventions instead [1611.05092][1807.08354][2508.08046][2502.16750].

The guarantees in TOG systems are also conditional rather than unconditional. Polygonal tracking requires speed bounds or nonempty controlled invariant sets; mixed-guard deployment depends on critical-region geometry; sequential defense depends on sensing radii, speed ratio, and angular positioning; EZ-aware interception depends on safe-set tightening under actuator saturation; crime PDEs exhibit different regimes depending on thresholds such as \(\chi_0\) and \(\chi_*\); and AI guardians achieve high but imperfect detection, with performance degrading under long attacks. This suggests that TOG research is centrally concerned with computable thresholds—speed ratios, bifurcation values, observability conditions, and detection rates—at which guardianship changes regime [2212.06628][2603.23649][2509.18146][2502.16750].

A final comparative feature is that the guardian can either stabilize or complicate system behavior, depending on the formulation. In polygonal environments, adaptive activation restores coverage as the offender becomes faster. In crime models, policing responsiveness can stabilize spatially uniform equilibria, but threshold crossings can also generate stationary hotspots, oscillatory cycles, or chaotic dynamics. In agentic systems, adding guardian agents improves detection yet does not eliminate vulnerabilities under persistent many-shot attack. The TOG framework is therefore best understood as a rigorous language for structured adversarial interaction, one in which the effectiveness of guardianship is analyzable but rarely monotone across all parameter regimes [1807.08354][2509.18146][2502.16750].

Source: https://www.emergentmind.com/topics/target-offender-guardian-system