AI-Security (PROTIVAI)

AI systems are attack surfaces. Treat them like it.

Organizations are deploying AI rapidly — LLM integrations, ML pipelines, model-serving infrastructure. But AI systems have unique attack vectors that traditional security tools don't understand: prompt injection, model poisoning, training data extraction, inference manipulation.

SPNT's AI-security module (PROTIVAI) secures your AI infrastructure. It monitors model registries, training pipelines, and inference endpoints for AI-specific threats — and writes findings to the same substrate that powers the rest of the platform.

What PROTIVAI monitors

• Model registries. MLflow, SageMaker Model Registry, Vertex AI Model Registry. Tracks model provenance, detects unauthorized model changes, validates model signatures.
• Training pipelines. Monitors training data sources, detects data poisoning attempts, validates training job integrity.
• Inference endpoints. API gateways, model serving infrastructure. Detects prompt injection attempts, adversarial inputs, inference abuse patterns.
• LLM integrations. ChatGPT, Claude, and custom LLM deployments. Monitors for data leakage, prompt injection, and unauthorized model access.

AI-specific threat detection

PROTIVAI detects attack patterns that traditional security tools miss:

• Prompt injection. Attempts to manipulate LLM behavior through crafted inputs. Direct injection, indirect injection, jailbreak attempts.
• Model poisoning. Malicious modifications to training data or model weights that alter model behavior in attacker-controlled ways.
• Training data extraction. Attempts to extract sensitive training data through model outputs — membership inference, model inversion attacks.
• Model theft. Attempts to extract model weights or replicate model behavior through excessive querying.
• Inference manipulation. Adversarial inputs designed to cause model misclassification or unexpected behavior.

AI supply chain security

PROTIVAI tracks the provenance of models and training data. Where did this model come from? What data was it trained on? Has it been modified since deployment?

For third-party models (Hugging Face, model marketplaces), PROTIVAI validates model integrity and monitors for known vulnerabilities in model architectures.

Integration with other modules

AI security findings flow through the same substrate as everything else:

• ISTRAGA reasons about AI threats in context of your broader security posture.
• REGULATIVA maps AI security findings to emerging AI governance frameworks (EU AI Act, NIST AI RMF).
• NADZOR tracks service account and API key access to AI infrastructure.