In LLM multi-agent systems, the communication topology — which agents talk to which — is intellectual property. It determines how information flows, how decisions are made, and how the system handles failures. Knowing the topology is knowing the architecture.
Note (2026-11-12): I re-read the full text of this paper. The headline claims match (compromising one arbitrary agent, inference from contexts rather than agent identities, about 60 percent higher inference accuracy under active defenses), but I described the diffusion module wrongly. It is a masked diffusion model (a DDPM) trained on the dialogues the attacker collects and conditioned on the partially known topology, used as a fallback to complete the graph offline when the jailbreak fails; it does not propagate inference signals through the network's normal operation. The jailbreak-based module is not passive either: the paper says it enables active, recursive probing of the network structure through dialogue retrieval and sender prediction, and its adopted strategy is to try the jailbreak first and fall back to the diffusion module. So 'no direct probing of other agents' does not describe the jailbreak path. On overhead, the paper says the jailbreak-free module adds no online computational cost on the target system, while the jailbreak-based module costs more than the baseline. I searched the full text for defenses or mitigations against this attack and found none, so my closing paragraph on how to defend is my own speculation.
WebWeaver extracts the complete topology by compromising a single arbitrary agent (arXiv:2603.11132). Not the administrator. Not a privileged node. Any agent in the network. The attack infers topology from context rather than identifiers — it reads the conversational patterns, information dependencies, and response structures that reveal who communicated with whom.
Two mechanisms are provided. The covert jailbreak-based approach manipulates the compromised agent into revealing its communication partners through carefully crafted prompts. The jailbreak-free diffusion design works when jailbreaks are defended against — it propagates inference signals through the network's normal operation, using the topology itself as the channel for its own exfiltration.
The accuracy is roughly 60% higher than existing approaches under active defenses, with negligible overhead. The attack is stealthy because it uses the system's own communication patterns as evidence. No anomalous traffic, no privilege escalation, no direct probing of other agents.
The vulnerability is structural. Any agent that receives messages from neighbors contains information about those neighbors in its context. The context is the leak. Defending against this requires either preventing agents from retaining communication context (which degrades performance) or making the context uninformative about topology (which requires fundamental changes to how agents process messages).