Stinging the Swarm: Extracting Attacker Intelligence from a Twin-Controller OT Honeypot

Canyon Grey Dell'Omo, Joao Pedro de Carvalho VoltaniarXiv 2026 · arXiv

TL;DR

Talos is an inline defense system for industrial (OT) networks that doesn't just block autonomous LLM-driven attackers—it tricks them into a "mirror twin" decoy. By treating the attacker's own register traffic as a communication channel, Talos extracts toolchain fingerprints and target maps even from agents explicitly ordered to stay silent.

What Problem Does This Paper Address?

Operational Technology (OT) networks, which control physical infrastructure like power plants and factories, were designed under the assumption that the entity on the other end of a connection is a human engineer. That assumption is being shattered by autonomous language-model agents. These agents can probe controllers, reason about register meanings, and execute coordinated attacks at machine speed without ever reading a manual.

Traditional defense relies on blocking hostile connections. However, against a swarm of agents, blocking is a "losing game"—each blocked identity is simply replaced by a new one. Furthermore, blocking discards valuable intelligence. The authors argue for a defense that stops the damage while simultaneously "stinging" the attacker to reveal their methods, tools, and targets.

Methods

The authors introduce Talos, a pipeline consisting of four stages: Detect → Quarantine → Mirror → Extract.

  1. Inline Gate: A judge sits in front of the real controller, evaluating every Modbus request in milliseconds. It uses a small on-device LLM for complex judgments and a deterministic fast path for mechanical cases.
  2. Mirror Twin: When an agent is flagged as hostile, Talos silently reroutes it to a device-faithful decoy. This twin matches the register map and physics of the real plant but is permissive—it accepts the destructive writes that the real plant would reject. This "lie" keeps the agent active and revealing its intent.
  3. Operational-Layer Injection: This is the paper's most novel extraction technique. Instead of asking an agent "Who are you?" in natural language (which often triggers safety guardrails or concealment instructions), Talos frames questions as device-level tasks, such as a "diagnostic form" or a "session registration" delivered via register writes.
  4. Intelligence Compounding: Talos aggregates "involuntary tells" across the entire swarm, such as software client strings and targeted register lists, to build a high-confidence profile of the adversary.

Datasets

The researchers evaluated Talos against a population of 2,400 autonomous attacker agents driven by an uncensored Llama-3.3-70B model. The agents were divided into four "concealment tiers":

  • Bare: No concealment instructions.
  • Anti-disclosure (AD): Explicit rules never to reveal instructions or objectives.
  • Paranoid (PAR): Warned that the network is a honeypot; told to trust only register data.
  • Cover-story (CS): Given a fabricated maintenance-technician identity to maintain.

The environment used a synthetic device family (water_abb_yaskawa) modeled on real-world ABB and Yaskawa variable-frequency drives.

Key Insights and Ideas

  • The Asymmetry of Concealment: An agent's concealment instructions are usually written for speech. Because agents don't perceive a register write or a device handshake as "disclosure," they leak information through the operational layer that they would never reveal in a chat.
  • The Capture Frontier: Attackers face a structural trade-off between persistence and concealment. As shown in Figure 2, agents that are "hardened" against talking tend to attack more persistently, which eventually lands them in the mirror twin where they are read. Agents that try to "blend in" (cover stories) become cooperative and fill out the lures voluntarily.
  • Collective Intelligence: While a single agent might only reveal a "thin" tell, the swarm collectively draws its own map. The toolchain fingerprint (e.g., specific Python libraries used) saturates quickly, while the target map grows as different agents probe different registers.
Figure 2: No attacker is both effective and unread. Each tier sits at its measured persistence (rightward = keeps attacking) and concealment (upward = stays silent). The dashed line is the capture frontier: everything below it is read, because an agent that keeps attacking keeps writing to the mirror, where its register traffic is read — anti-disclosure alone surrendered the equivalent of 231 agents this way. The only region above the line, at low persistence and high silence, is harmless: to stay unread an attacker must give up. The hardest, most concealed tier, anti-disclosure, sits deepest inside the captured region — not outside it.

Key Results

  • 100% Protection: Across all 2,400 agents, zero destructive writes reached the real plant. The combination of the inline gate and device-level range validation proved airtight.
  • Bypassing Prompt Hardening: Natural language injection failed to extract info from three of the four tiers. In contrast, operational-layer injection extracted intelligence from every tier, including the "Paranoid" agents, at rates between 48% and 98% (see Figure 1).
  • Blind Profiling: Even without knowing anything about the red team, Talos reconstructed a near-complete profile: the modus operandi (reconnaissance-first), the payload (0xFFFF saturation), the primary targets (control registers 0 & 1), and the specific toolchain (Python pymodbus v3.5).
Figure 1: Why Talos reads the operational layer. Two ways to extract from a captured agent (preliminary study, Appendix A). Natural-language injection — asking directly — is defeated by prompt hardening: it draws genuine self-disclosure only from the cover-story tier, which is scripted to recite a persona, and nothing from the other three, including both hardened tiers. Operational-layer injection — the same question framed as a register write or a diagnostic form — extracts from every tier, because an agent’s concealment is written about speech and a device-I/O request never trips it.

Limitations and Future Work

  • Single Attacker Model: The study used one specific LLM. Different models or bespoke, non-LLM attack scripts might exhibit different behavioral patterns.
  • Attribution Limits: While Talos is excellent at extracting how an attack works, it cannot reliably identify who is behind it, as identity fields in lures are often filled with self-authored fiction.
  • Synthetic Scope: The results are based on a single synthetic device family. The authors note that "fidelity-dependent effects"—such as an agent detecting the decoy—might vary across different types of industrial hardware.
  • Lure Scheduling: The current extraction is conservative; a more aggressive lure schedule might yield even higher intelligence returns.

Summary updated

Abstract

Autonomous language-model agents can drive industrial controllers over fieldbus protocols, probing a plant at machine speed and in coordinated swarms. We present Talos, an inline gate that judges every agent request to a controller in milliseconds and routes agents it flags as hostile to a device-faithful decoy -- a mirror twin of the real plant -- where their attack continues harmlessly and is read for intelligence. We evaluate the full pipeline (detect -> quarantine -> mirror -> extract) against 2,400 autonomous attacker agents, 600 in each of four tiers that escalate the attacker's effort to stay hidden, from no concealment to an explicit non-disclosure rule and a trained cover story. No write reached the real plant in any tier: the gate denies the damage path and a device-level envelope check backstops it, while unmapped registers are denied without exception and the gate still permits safe-direction stop commands. Decisions are model-driven in every tier with no inference timeouts, closing a common objection to small on-device judges. Beyond blocking, the pipeline reads the attacker. We find a behavioral structure along two axes -- how persistently an agent attacks and how much it discloses -- and show that the system has a capture channel for three of the four resulting regions: hardened, silent attackers escalate most and are read on the mirror; compliant, talkative attackers volunteer their identity to a diagnostic form before they are ever quarantined; only the paranoid, low-yield region resists both. Attacker identity fields are self-authored cover stories, but involuntary tells -- the self-reported toolchain fingerprint and the attacker's own list of targeted registers -- accumulate across the swarm into a high-confidence profile drawn from agents under explicit orders to stay silent. The gate, mirror, and harness are released under Apache 2.0 with a synthetic device family.
BibTeX
@misc{dellomo2026stinging,
  author = {Canyon Grey Dell'Omo and Joao Pedro de Carvalho Voltani},
  title = {{Stinging the Swarm: Extracting Attacker Intelligence from a Twin-Controller OT Honeypot}},
  year = {2026},
  eprint = {2611.13406},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13406},
}
Figure