Summary
This proposal suggests adding a deterministic, physics-based Anomaly Detection (AD) module as a new MCP Server within the AssetOpsBench ecosystem. While the current TSFM server excels at statistical anomaly detection (catching deviations from learned norms), it is vulnerable to "slow drift" attacks (e.g., Stuxnet-style centrifuge manipulation) where the model gradually learns the malicious state as the new normal. A physics-based anchor provides an un-spoofable ground truth.
Motivation
In Industrial Control Systems (ICS) and cyber-physical security, attackers do not need to break the math of the ML model; they just need to lie to the sensors. If a flow sensor reports normal pressure while a valve is physically closed, statistical forecasting might not flag it if the attacker ramps the fake data slowly enough.
Newtonian physics, however, cannot be spoofed. If mass balance (mass in = mass out) or thermodynamic laws (energy in = work + heat) are violated, a cyber-physical attack is occurring regardless of what the statistical model thinks.
Proposed Architecture
We propose a Physics Validation MCP Server (similar to the existing Vibration or FMSR servers) that exposes tools like:
validate_mass_balance(asset_id, inlet_flow, outlet_flow)
validate_thermodynamics(asset_id, power_draw, heat_dissipation)
Instead of returning a probability score, this server returns a strict boolean: physics_violation: True/False.
Proof of Concept
I have built an open-source prototype of this architecture called the Cyber-Kinetic Defense Grid (CKDG). It uses a Rust-based physics compiler to validate telemetry streams against hard physical invariants. When a violation occurs, it triggers an LLM agent to explain the attack vector to the operator.
Repo: https://github.com/Ijas14/cyber-kinetic-defense-grid
Integration Path
If the AssetOpsBench team is interested, the next step would be defining an ICS security scenario (e.g., SWaT dataset water treatment spoofing) where the agent must use both the TSFM server (for statistical anomalies) and the proposed Physics server (for hard physical violations) to successfully diagnose the attack.
I would love to hear the team's thoughts on whether deterministic physics anchors fit into the benchmark's future roadmap.
Summary
This proposal suggests adding a deterministic, physics-based Anomaly Detection (AD) module as a new MCP Server within the AssetOpsBench ecosystem. While the current TSFM server excels at statistical anomaly detection (catching deviations from learned norms), it is vulnerable to "slow drift" attacks (e.g., Stuxnet-style centrifuge manipulation) where the model gradually learns the malicious state as the new normal. A physics-based anchor provides an un-spoofable ground truth.
Motivation
In Industrial Control Systems (ICS) and cyber-physical security, attackers do not need to break the math of the ML model; they just need to lie to the sensors. If a flow sensor reports normal pressure while a valve is physically closed, statistical forecasting might not flag it if the attacker ramps the fake data slowly enough.
Newtonian physics, however, cannot be spoofed. If mass balance (mass in = mass out) or thermodynamic laws (energy in = work + heat) are violated, a cyber-physical attack is occurring regardless of what the statistical model thinks.
Proposed Architecture
We propose a Physics Validation MCP Server (similar to the existing Vibration or FMSR servers) that exposes tools like:
validate_mass_balance(asset_id, inlet_flow, outlet_flow)
validate_thermodynamics(asset_id, power_draw, heat_dissipation)
Instead of returning a probability score, this server returns a strict boolean: physics_violation: True/False.
Proof of Concept
I have built an open-source prototype of this architecture called the Cyber-Kinetic Defense Grid (CKDG). It uses a Rust-based physics compiler to validate telemetry streams against hard physical invariants. When a violation occurs, it triggers an LLM agent to explain the attack vector to the operator.
Repo: https://github.com/Ijas14/cyber-kinetic-defense-grid
Integration Path
If the AssetOpsBench team is interested, the next step would be defining an ICS security scenario (e.g., SWaT dataset water treatment spoofing) where the agent must use both the TSFM server (for statistical anomalies) and the proposed Physics server (for hard physical violations) to successfully diagnose the attack.
I would love to hear the team's thoughts on whether deterministic physics anchors fit into the benchmark's future roadmap.