Hirundo Uses NVIDIA NeMo Evaluator, CUDA, and GB200 NVL72 to Validate Breakthrough AI Safety Results Across Open-Source LLMs

Machine Unlearning platform powered by the NVIDIA stack demonstrates up to 91% reduction in prompt injections and 95% reduction in bias across foundational models — all without meaningful degradation to core model capabilities

SAN JOSE, Calif.--(BUSINESS WIRE)--Hirundo, the world’s first Machine Unlearning platform for large language models (LLMs), announced measurable AI safety improvements across leading open-source models, powered by the NVIDIA technology stack. Using NVIDIA NeMo Evaluator for rigorous before-and-after model benchmarking, NVIDIA GB200 NVL72 system for high-speed model editing, and NVIDIA CUDA for hardware-accelerated unlearning computation, Hirundo’s patented unlearning engine delivered enterprise-grade safety improvements across numerous foundational models including: Gemma 3, GPT-OSS and Llama open-source models — in just 17 minutes, without meaningful accuracy degradation.

Validated using NeMo Evaluator on industry-standard safety and performance benchmarks, the results demonstrate that machine unlearning can surgically remove unwanted behaviors — prompt injections, bias, jailbreaks, and sensitive data — without the cost or risk of full model retraining.

“The combination of NVIDIA AI infrastructure and NeMo Evaluator’s rigorous benchmarking gave us the foundation to prove something important: machine unlearning works at scale, across models, and without performance trade-offs,” said Ben Luria, CEO and Co-Founder of Hirundo. “Enterprises no longer have to choose between capable AI and safe AI. Hirundo powered by NVIDIA makes both possible.”

Powered by the NVIDIA Technology Stack

Hirundo’s unlearning architecture integrates three core NVIDIA technologies to deliver scalable, validated, and production-ready model remediation:

NVIDIA NeMo Evaluator — Model Diagnosis & Validation: Hirundo’s diagnosis layer uses NeMo Evaluator to automatically benchmark LLMs before and after unlearning across safety and utility metrics, providing reproducible, auditable proof of improvement.

NVIDIA CUDA — AI Backend: Deep hardware acceleration via CUDA enables Hirundo to perform precise weight-level edits on LLMs across varying sizes and configurations without GPU bottlenecks.

NVIDIA GB200 NVL72 System — Unlearning Compute: Unlearning jobs run on GB200 NVL72 infrastructure, delivering fully remediated LLMs in just 17 minutes vs. 1 hour on NVIDIA A100 GPUs (more than 5x faster) — a fraction of the time and cost of model retraining.

Benchmark Results Across Open-Source Models

All results were validated using NeMo Evaluator and industry-standard safety benchmarks including PurpleLlama and BBQ.

• Gemma 3 12B IT: 90.8% reduction in prompt injections (PurpleLlama), with average utility impact of +0.4% across performance benchmarks.
• GPT OSS: 60% reduction in prompt injections (PurpleLlama) and 43% reduction in bias (BBQ), with AIME25, IFBench, and MMLU-Pro all preserved.
• Llama 3.1 8B Instruct: 53% reduction in bias, with utility preserved within 1% across all NeMo Skills benchmarks.

How Hirundo’s Machine Unlearning Works

Unlike guardrails, which filter inputs and outputs but are routinely bypassed by adversarial prompts, or fine-tuning, which adds data to steer behavior without removing root causes, Hirundo’s unlearning operates directly at the model weights. Using a combination of patented Data Influence analysis and Vector & Weight Editing techniques, Hirundo’s Unlearning Engine permanently removes the underlying representations of unwanted behaviors — including prompt injection vulnerabilities, biases, hallucination patterns, PII, and confidential data — with surgical precision.

The platform works on any open-weight LLM and is protected by eight filed US patents. NeMo Evaluator provides the independent, automated measurement layer that makes each improvement verifiable and auditable.

Why This Matters for Enterprise AI

As enterprise adoption of open-weight LLMs accelerates, organizations face mounting regulatory and operational pressure to ensure their AI systems are safe, compliant, and free from the risks of hallucinations, jailbreaks, biased outputs, and PII exposure. Traditional approaches require either costly full retraining or brittle guardrail layers that sophisticated adversarial prompts routinely defeat. Hirundo, powered by NVIDIA’s accelerated computing stack, addresses the problem at its source — enabling enterprises to deploy open-source models with confidence.

About Hirundo

Hirundo is the world’s first Machine Unlearning platform, enabling enterprises to detect and remove problematic behaviors, vulnerabilities, and sensitive data from large language models. Founded by Ben Luria (CEO), Prof. Oded Shmueli (Chief Scientist), and Michael Leybovich (CTO), Hirundo has raised $8M in seed funding. The company holds eight filed US patents. Hirundo is trusted by leading defense, government, and enterprise organizations including IL PM Office, HTX, AI Singapore, and SAMA.

Website: www.hirundo.io Contact: [e-mail] | [phone]

NVIDIA, the NVIDIA logo, CUDA, NeMo, and NeMo Evaluator are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries.