Stopping Data Poisoning of Grid Models

Maintenance, forecasting, and voltage models rely on sensor data, work orders, and documents that attackers can manipulate. Poisoned or injected inputs can cause models to misread conditions and issue unsafe recommendations that keyword filters never flag. PointGuard AI inspects inputs and tool outputs for injection and poisoning patterns, tests models against manipulated data, and alerts teams before recommendations reach operators.

Grid models trust data that adversaries can influence.

• Sensor feeds and documents can be manipulated upstream.
• Poisoned training data skews model behavior over time.
• Unsafe recommendations can look plausible to operators.
• Keyword filters miss subtle manipulation.

Runtime Guardrails, AI Red Teaming, and Supply Chain Security inspect inputs and test models against manipulation.

1. Inspect inputs. Scan retrieved documents, work orders, and tool outputs for injection patterns using fine-tuned small language models at sub-100 ms latency.

2. Test against poisoning. Probe models with manipulated data scenarios before release and after retraining.

3. Verify model sources. Score models and datasets for provenance, tampering, and unsafe formats before deployment.

4. Alert and escalate. Flag anomalous recommendations for operator review and map findings to OWASP for reporting.

• OWASP Top 10 for LLMs: LLM01 Prompt Injection, LLM04 Data and Model Poisoning, and LLM09 Misinformation, covering injected inputs, poisoned data, and unsafe recommendations.
• OWASP Top 10 for Agentic Applications: ASI06 Memory and Context Poisoning, addressing manipulated memory and context that skew agent decisions.
• NIST AI RMF: MAP identifies data sources; MEASURE tests resilience; MANAGE blocks manipulated inputs; GOVERN sets data integrity standards.