Security Policy & Threat Model

Security Policy & Threat Model

Industrial IoT edge nodes operate in adversarial or exposed network environments (factory floors, substation gateways, remote pump stations). Sofia Engine is built around a comprehensive threat model and a defense-in-depth security architecture.

---

1. Trust Boundaries

Sofia defines six explicit trust boundaries (documented in specs/001-sofia-engine-modernization/threat-model.md):

[ Field Network / PLCs / Sensors ]
              |
====== TB-1: Transport Boundary ======
              |
[ Telemetry Adapters (MQTT, Modbus, Serial) ]
              |
      [ Bounded Ring Buffers ]
              |
[ DSP & Feature Extraction Core ]
              |
====== TB-2: Model Supply Boundary ======
              |
[ Model Manifests & Inference Backends ]
              |
[ Diagnostic Engine & Evidence Scoring ]
              |
====== TB-4: Operator Boundary ======
              |
[ CLI / API / Copilot ]
              |
====== TB-5: Actuation Boundary ======
              |
[ PolicyEngine (Default DENY) ]
              |
[ Physical Actuators / PLCs ]

---

2. Core Security Controls

1. Absolute Ban on Insecure Deserialization (SOFIA-SEC-001)

Python's native pickle module allows arbitrary code execution during deserialization. Sofia Engine enforces:

• Zero pickle across the entire codebase: Verified via automated AST scans in tests/security/test_no_pickle.py.
• Telemetry, features, and configurations are serialized exclusively via strict JSON, JSONL, or NumPy .npz binary formats.

2. Model Supply Chain Integrity (SOFIA-SEC-002 & 003)

All machine learning models and anomaly detector configurations require a validated ModelManifest:

• Schema Validation: Explicit typing of input feature names, tensor shapes, and preprocessing version.
• Cryptographic Checksum: SHA-256 hash of the artifact file must match the manifest. Mismatched artifacts are rejected loudly.
• Digital Signatures: Optional Ed25519 asymmetric signature verification for air-gapped deployments.

3. Replay & Injection Defense (SOFIA-SEC-004 & 006)

• Monotonic Sequence Numbers: Telemetry samples carry sequential IDs; duplicates and backwards timestamps are flagged DUPLICATE or STALE.
• Payload Ceilings: Ingest buffers and network adapters reject frames exceeding declared byte limits (default 1 MiB) to prevent memory exhaustion attacks.
• Strict Schema Parsing: CSV and JSONL parsers reject newline or delimiter injection attempts.

4. Default "DENY" Control Path (SOFIA-SAFE-001)

The PolicyEngine ensures that inference models cannot directly trigger equipment movements. Every command requires operator approval, equipment operational state verification, rate limiting, and physical interlock clearance.

---

3. Threat Model Enumeration (Selected Highlights)

Threat ID	Description	Impact	Implemented Mitigation
T-01	Telemetry Value Poisoning	Skew anomaly detectors to mask faults or trigger false trips.	Robust estimators (MAD, median), physical range limits, and quality marking OUT_OF_RANGE.
T-02	Malformed Protocol Frames	Buffer overflows or parser crash in Modbus/Serial streams.	Strict frame length validation, typed protocol exceptions, drop-and-count strategy.
T-03	Protocol Injection	Control character injection into CSV/JSONL streams.	Fixed column mappings, json.loads only, rejection of unescaped newlines.
T-04	Telemetry Replay Attack	Replaying old healthy sensor streams to disguise failing hardware.	Monotonic sequence validation, TTL freshness window, clock anomaly detection.
T-06	Compromised MQTT Broker	Malicious broker floods topic with oversized payloads.	Strict TLS encryption, payload size limits (1 MiB), memory-bounded ring buffers.
T-11	Model Artifact Tampering	Replacing model weights with adversarial backdoor.	SHA-256 hash enforcement in manifest, optional Ed25519 signature checks.
T-17	Actuation Bypass	Attacker triggers PLC movements bypassing safety rules.	Universal PolicyEngine gate; no bypass path exists in library code.

---

4. Vulnerability Disclosure Policy

If you discover a security vulnerability in Sofia Engine:

• Do NOT open a public GitHub issue.
• Open a private security advisory on GitHub: Security $\to$ Report a vulnerability.
• Or contact the maintainers directly via Rootcastle Engineering & Innovation at security@rootcastle.com (or as outlined in SECURITY.md).

Rootcastle acknowledges reports within 5 business days and provides coordinated remediation timelines.