A Hybrid Causal-Generative Multi-Agent Framework for Secure Autonomous Mechatronics: Mitigating Real-Time Prompt Injections in Edge-Deployed Disaster Recovery Networks

As large language models increasingly assume command of autonomous mechatronic swarms in disaster recovery operations, a structurally underexplored vulnerability emerges: realtime prompt injection delivered through legitimate sensory channels. Unlike conventional cyber intrusions, these attacks subvert the semantic reasoning layer itself, translating adversarial natural language inputs into irreversible kinetic consequences through motor controllers and actuators. Existing defenses rely on cloud-based API guardrails or secondary monitoring models that are architecturally incompatible with edge-deployed, communication-denied environments. This paper presents the Hybrid Causal-Generative Multi-Agent Framework (HCG-MAF), which interposes a lightweight Causal Safety Enforcer (CSE) between the LLM output stage and the microcontroller interface. The CSE maintains a dynamically calibrated Structural Causal Model (SCM) of each agent's physical environment and applies Pearl's do-calculus to evaluate the counterfactual kinetic consequences of every generated directive before actuation. Across 400 adversarial trials spanning three injection vectors, drawn from a wider 772-sample Sim-to-Real cohort, HCG-MAF achieves a 98.5% Intervention Success Rate (ISR, 394/400) at a 42 ms mean safety-decision latency with a 1.2% false-positive rate, all without cloud connectivity; on transient cases inside the first 45 ms the ISR falls to 96.8%. A Byzantineresilient gossip-based consensus protocol secured through cryptographic causal attestation maintains swarm coherence under simultaneous compromise of up to one-third of deployed agents, provided the n ≥ 3f + 1 bound is satisfied.

No artifact released. This paper reports measurements, but its code, corpus and raw outputs are not public, so the numbers above cannot be independently checked. Treat the results as unreproduced.

@misc{pandey2026hybrid,
  title  = {A Hybrid Causal-Generative Multi-Agent Framework for Secure Autonomous Mechatronics: Mitigating Real-Time Prompt Injections in Edge-Deployed Disaster Recovery Networks},
  author = {Aashika Pandey and Sushant Poudel},
  year   = {2026},
  note   = {AIMS 2026 — 23–24 July 2026; IEEE Xplore},
  url    = {https://sushantpoudel2028.com.np/papers/hcg.pdf}
}

Claims about his own work that the ledger covers are re-checked weekly against their public source by sushant-me/reputation — 27 claims today, 19 checked live. Coverage is stated rather than implied: it does not yet extend to every manuscript listed here. The reproducibility note above is part of that record: where no artifact exists, the page says so instead of implying one.