Video by H2O.ai via YouTube
How Enterprise h2oGPTe protects LLM applications from toxic content, PII leaks, and adversarial jailbreak attempts.
Even high-performing generative AI models require safeguards. h2oGPTe enforces multi-stage guardrails at the collection level—monitoring content during ingestion, at prompt submission, and before final response generation. Built-in toxic topic classifications and configurable custom guardrails keep AI strictly on-topic. PII detection uses a defense-in-depth approach combining regex, Presidio, and a fine-tuned ModernBERT model, while PromptGuard actively blocks adversarial jailbreak patterns and logs every violation.
Technical Capabilities & Resources
➤ Toxic Content & Custom Topic Filtering: Block harmful content and restrict AI to approved business topics using configurable guardrails.
🔗 https://docs.h2o.ai/enterprise-h2ogpte/guide/collections/create-a-collection#guardrails-and-pii-detection
➤ PII Detection & Redaction: Identify and redact sensitive data across prompts and responses using Regex, Presidio, and ModernBERT.
🔗 https://docs.h2o.ai/enterprise-h2ogpte/guide/collections/pii-sanitization#pii-detection-methods
➤ Adversarial Jailbreak Protection: PromptGuard detects and neutralizes adversarial prompt patterns before they reach the model.
🔗 https://docs.h2o.ai/enterprise-h2ogpte/changelog/tags/v-1-5#guardrails