When we reason about the security of a networked system—whether it is a power grid, a social network, or a collective of Large Language Model (LLM) agents—our first instinct is often to count the number of compromised nodes. We assume that a system with fewer malicious actors is inherently safer than one with many.
However, structural network analysis reveals a more nuanced reality: where an adversary strikes matters far more than how hard they strike. The structural topology of a network dictates how easily an attack propagates, how effectively a defense can be deployed, and whether a system will gracefully degrade or catastrophically collapse under duress.
Centrality as an Amplifier and a Shield
To understand why topology governs robustness, we must look at how information flows through structural bottlenecks. In any non-uniform network, certain nodes hold disproportionate structural power due to their position.
Xiaotong Cheng et al. demonstrated this phenomenon in the context of information cascades in Online Influence Maximization with Semi-Bandit Feedback under Corruptions. They found that a single malicious agent positioned at a high-centrality bottleneck can disrupt an entire network's information flow more effectively than a large crowd of adversarial nodes situated at the periphery. When an adversary targets these high-load bridges, they exploit the network's reliance on efficient paths. As shown by K.-I. Goh et al. in Classification of scale-free networks, tree-like scale-free networks are incredibly fragile because removing these central hubs completely fragments the graph, cutting off communication between remaining clusters.
Peripheral Attack (Low Impact) Central Bottleneck Attack (High Impact)
O O O O
\ / \ /
X ---- O ---- O O ---- X ---- O
/ \ / \
O O O O
(Only one node isolated) (Entire network fragmented)
Conversely, this structural asymmetry can be weaponized by defenders. Instead of trying to secure every node equally, defenders can focus resources on these exact same bottlenecks. Haodi Zhang et al. applied this intuition to cybersecurity in A Method for Deploying Distributed Denial of Service Attack Defense Strategies on Edge Servers Using Reinforcement Learning. By mapping the upstream and downstream relationships of edge servers, their framework identifies strategic "choke points" to deploy traffic filters. Intercepting malicious traffic at these topological bottlenecks prevents individual nodes from being overwhelmed and stops the attack before it reaches the core network.
The Connectivity Paradox
If central hubs create vulnerabilities, is the solution simply to connect everyone to everyone else? Intuition suggests that high connectivity adds redundancy, making a network more robust. In reality, connectivity acts as a double-edged sword.
In routing and consensus protocols, high connectivity is a powerful defense. Martin Byrenheid et al. showed in On the Limits of Byzantine-tolerant Spanning Tree Construction in Route-Restricted Overlay Networks that high local clustering provides natural resilience against denial-of-service attacks. If a malicious node tries to disrupt a spanning tree, highly connected neighbors can easily route around the failure using alternative honest paths.
However, when a system is exposed to a contagious threat—like malware or misinformation—high connectivity becomes a liability. In G-Safeguard: A Topology-Guided Security Lens and Treatment on LLM-based Multi-agent Systems, Kun Wang et al. modeled adversarial prompt injections in multi-agent LLM systems as a spreading virus. They discovered that highly connected topologies, such as star or random graphs, allow a single compromised agent to corrupt the entire system within a few interaction rounds. Similarly, Rana Muhammad Shahroz et al. demonstrated in Agents Under Siege: Breaking Pragmatic Multi-Agent LLM Systems with Optimized Prompt Attacks that path redundancy gives attackers more opportunities to find and exploit sequences of agents with weaker safety constraints.
This tension creates a fascinating contrast depending on the attack vector:
- Message Manipulation Attacks: Linear chain structures are highly vulnerable. As Pengfei He et al. showed in Red-Teaming LLM Multi-Agent Systems via Communication Attacks, an "Agent-in-the-Middle" can easily manipulate messages in a line because there are no alternative paths for agents to cross-verify information. Highly connected graphs are more resilient here because agents can challenge inconsistent data.
- Contagious Jailbreak Attacks: Linear chains act as natural filters. Tianyi Men et al. found in A Troublemaker with Contagious Jailbreak Makes Chaos in Honest Towns that because an adversarial prompt must traverse multiple communication hops—each with independent safety filters and memory resets—the jailbreak attempt often degrades and fails before reaching the rest of the network.
This non-linear relationship means that the most vulnerable networks are often those with moderate connectivity. Yigit Ege Bayiz et al. formalized this in The Effect of Network Topology on the Equilibria of Influence-Opinion Games, proving that moderately connected networks allow adversaries to easily isolate groups into echo chambers, whereas extreme local or global connectivity naturally restricts an attacker's control.
Weaponizing and Defending Topological Information
Because topology is so decisive, an attack's success depends heavily on the adversary's structural awareness. If an attacker has perfect knowledge of a network, they can execute highly efficient, targeted strikes on high-degree hubs or critical structural bridges (Asim Kumer Dey et al., Motif-based analysis of power grid robustness under attacks).
This dependency opens up a novel avenue for defense: topological deception.
Instead of reinforcing physical infrastructure, defenders can manipulate the information the attacker receives. Jun Wu et al. explored this in Enhancing structural robustness of scale-free networks by information disturbance. By introducing slight perturbations to the reported node degrees—effectively making low-traffic nodes look like critical hubs—they showed that an attacker wasting resources on these false targets can nearly double the structural robustness of the network.
True Network Topology Adversary's Disturbed View
O (Low Degree) (Hub)
/ / | \
O --- Hub --- O O | O
\ |
O (Low Degree) (Hub)
When attackers realize they are being monitored, the situation evolves into a strategic game of cat-and-mouse. In Hiding From Centrality Measures: A Stackelberg Game Perspective, Marcin Waniek et al. analyzed how covert actors alter their connections to evade detection. They noted that because no single centrality metric is universally optimal across all graph types, an evader can successfully hide from path-based metrics while remaining completely visible to degree-based metrics.
Ultimately, understanding topology-dependent robustness shifts our perspective from static defense to dynamic, structural adaptation. Whether by pruning edges to quarantine a multi-agent system or altering physical parameters to disrupt an attacker's feedback loops in a power grid (Subhash Lakshminarayana et al., Moving-Target Defense for Detecting Coordinated Cyber-Physical Attacks in Power Grids), the most resilient systems are not those that are built to be impenetrable, but those designed to manipulate the structural geometry of the attack itself.