Energy utilities manage power grids, pipelines, and generation facilities that serve millions. AI can optimize load balancing, predict equipment failures, and detect cyber threats. The problem: a compromised AI agent connected to operational technology (OT) systems could cause physical damage to infrastructure or disrupt power delivery to entire regions.
AKIOS solves this with the Security Cage: an ephemeral, sandboxed runtime where AI analyzes grid data under strict, code-defined policies — with no direct access to control systems.
The Problem
The energy sector's challenge is unique: unlike banking or healthcare, a breach doesn't just expose data — it can cause physical harm. A compromised AI with write access to SCADA systems could open breakers, reroute power, or mask equipment failures. Traditional cybersecurity tools were designed for IT networks, not for the OT/IT convergence that modern smart grids demand.
AKIOS enforces an absolute boundary between AI analysis and operational control.
The Regulatory Landscape
US energy companies face some of the strictest infrastructure protection requirements:
| Regulation | Scope | How AKIOS Enforces It |
|---|---|---|
| NERC CIP | Mandatory cybersecurity standards for bulk electric systems — access controls, electronic security perimeters, incident reporting | Sandbox satisfies electronic security perimeter requirements. Every data access logged, every output signed. |
| FERC Orders | Supply chain risk management and cybersecurity incident reporting for automated systems | AI runs in ephemeral containers with no supply chain dependencies. Zero third-party data transmission. |
| DOE Cybersecurity Guidelines | Air-gapped architectures, human-in-the-loop controls, and data integrity for AI/ML in energy | Complete network isolation from OT networks. AI can recommend but never execute control commands. |
| TSA Pipeline Security | Mandatory cybersecurity measures including network segmentation for pipeline operators | Policy-enforced network segmentation — pipeline telemetry enters read-only, no write path exists. |
| ICS-CERT Advisories | CISA advisories on industrial control system vulnerabilities and AI-specific threat vectors | Seccomp-bpf filters block all syscalls that could interact with control system protocols. |
AKIOS enforces these at the runtime level — the AI agent never operates outside the compliance boundary.
The Concept: Policy as Code
AKIOS introduces the concept of a "Security Cage" — an ephemeral, sandboxed runtime where data is processed under strict, code-defined policies. For energy, the key innovation is the absolute read-only boundary: the AI can read telemetry, but the cage is physically incapable of sending commands to control systems.
The Workflow: Grid Anomaly Detection
| Step | What Happens | Security Control |
|---|---|---|
| 1. Ingestion | SCADA telemetry and sensor readings loaded into the cage | Asset IDs, coordinates, and facility details abstracted. Read-only filesystem agent. |
| 2. Abstraction | Critical infrastructure identifiers replaced with generic tokens | Substation names, GPS coordinates, and asset serial numbers never reach the LLM. |
| 3. AI Analysis | LLM analyzes sensor patterns — load imbalances, degradation signatures, unusual access patterns | Budget capped ($0.75/scan), complete OT network isolation, no command execution capability. |
| 4. Alerting | Anomalies classified by severity and type with confidence scores | AI can recommend actions but cannot execute any commands on operational systems. |
| 5. Audit | Every inference and data access cryptographically signed into a Merkle chain | NERC auditors can verify the complete analysis path for any alert generated. |
Architecture
graph LR
SCADA["SCADA / OT\nControl Systems"] -->|"telemetry only\n(one-way diode)"| FS["filesystem agent\nread-only"]
subgraph CAGE["AKIOS Security Cage · seccomp-bpf"]
FS --> PII["Abstraction Engine\n«ASSET_ID» «COORD» «SERIAL»"]
PII --> LLM["llm agent\nanomaly detection"]
LLM --> TE["tool_executor\nseverity classification"]
TE --> VALID["Output Validation\nno control commands"]
VALID --> MERKLE["Merkle Chain\nSHA-256 signed"]
MERKLE --> COST["Cost Kill-Switch\n$0.75 / scan"]
end
COST -->|"anomaly alerts\n(abstracted)"| Alert["Alert Dashboard\nSOC / NOC"]
Alert --> Engineer["Grid\nEngineer"]
MERKLE -->|"audit export\n(immutable)"| NERC["NERC CIP\nAudit Trail"]
NERC --> FERC["FERC\nCompliance"]
Policy Configuration
The entire compliance posture is defined in a single YAML file:
# energy-nerc-cip-policy.yml
security:
sandbox: strict
network: isolated
allowed_endpoints: [] # zero network access — absolute OT/IT separation
pii_redaction:
enabled: true
patterns: [asset_id, gps_coordinates, serial_number, substation_name]
mode: aggressive
budget:
max_cost_per_run: 0.75
currency: USD
audit:
merkle_chain: true
export_format: jsonl
retention_days: 1095 # 3 years — NERC CIP retention requirement
constraints:
block_control_commands: true # hard-coded: AI cannot generate SCADA commands
read_only: true
What the Grid Engineer Sees
At the end of the workflow, the grid engineer receives a structured alert:
| Field | Value |
|---|---|
| Alert ID | GRID-2026-0210-****6721 |
| Severity | 🟠 High — Equipment degradation signature |
| Pattern Detected | Transformer harmonic distortion increasing 12% over 72 hours — consistent with insulation breakdown |
| Recommended Action | Schedule preventive maintenance inspection within 14 days |
| Confidence | 88% |
| Audit Hash | c9d4e7...b23a |
| Infrastructure Data Exposed | ❌ None — all asset identifiers abstracted before analysis |
No substation names. No GPS coordinates. No equipment serial numbers. Just actionable maintenance intelligence with a cryptographic proof chain.
Why It Matters
- OT/IT Separation: The Security Cage is completely air-gapped from operational technology networks. The AI reads telemetry — it can never send commands to control systems.
- Zero-Persistence Processing: Sensor data is processed in ephemeral containers destroyed after analysis. No grid data is ever stored, cached, or retained by the AI model.
- NERC CIP Compliance: The sandbox satisfies electronic security perimeter requirements. Every data access is logged, every output is signed, every session is containerized.
- Physical Safety: Hard-coded policy rules prevent the AI from generating or outputting anything resembling a control command. The boundary between analysis and action is absolute.
- Predictive Maintenance: AI can identify equipment degradation patterns weeks before failure, enabling proactive maintenance without exposing critical infrastructure details.
Try It Yourself
pip install akios
akios init my-project
akios run templates/file_analysis.yml
Secure your AI. Build with AKIOS.