- NVIDIA’s H100 and H200 GPUs enable confidential computing for private AI inference with less than 5% latency overhead.
- Approximately 96% of executives expect a generative AI security breach within three years, as current corporate security is not designed for AI systems.
- Privacy-enhancing technologies and comprehensive frameworks are essential for end-to-end AI security, protecting data across the full pipeline from collection to inference.
Roughly 96% of executives expect a security breach tied to generative AI within three years, according to the IBM Institute for Business Value. That figure reflects a specific problem: most corporate security was built for databases and networks, not for AI systems that can leak sensitive information through their own outputs.
A specific problem is that most corporate security was built for databases and networks, not for AI systems that can leak sensitive information through their own outputs.
Enterprise AI’s Urgent Need for Data Security
The risks are unlike anything traditional cybersecurity was designed to handle. Prompt injection attacks can manipulate an AI model into ignoring its instructions. Unintentional data leakage can occur when a model reproduces fragments of its training data in responses. Then there is “shadow AI”: employees using unauthorised tools that bypass corporate controls entirely. Regulatory pressure adds another layer, with GDPR, HIPAA and the EU AI Act all requiring organisations to show precisely how their AI systems handle sensitive information.
Confidential Computing Becomes the 2026 Standard
The most direct technical answer to these risks is confidential computing. The approach processes data inside hardware-enforced secure enclaves, so data stays encrypted even while the model is actively working with it. NVIDIA‘s H100 and H200 GPUs support this capability, creating isolated environments where model weights, training data and inference inputs are all protected. Research by Hazy found that confidential LLM inference on models like Qwen-32B, with inputs of around 8,000 tokens, can run with less than 5% latency overhead. That closes what some have called the “AI privacy paradox”: the idea that powerful AI and data confidentiality cannot coexist.
Hardware enclaves are not the whole picture. A set of privacy-enhancing technologies, known as PETs, adds further layers of defence. Fully homomorphic encryption (FHE) allows computations to run directly on encrypted data, so the raw information is never exposed during processing. Federated learning trains models on data that never leaves its source, which is useful for healthcare or financial data spread across multiple organisations. Differential privacy adds calibrated mathematical noise to training data, making it mathematically difficult to identify any individual from the model’s outputs. Together, these tools protect data across the full pipeline, from ingestion through to the final response.
Frameworks Redefine End-to-End AI Security
A comprehensive framework for securing generative AI typically covers data collection and handling, model development and training, live inference, and the underlying infrastructure, also including AI governance, monitoring for fairness, bias and model drift over time.
Other frameworks address the broader picture. NIST‘s AI Risk Management Framework and the OWASP Top 10 for LLM Applications each map specific vulnerabilities and mitigation strategies for AI systems. Microsoft’s Cloud Adoption Framework advises supplementing standard threat modelling with AI-specific risk catalogues such as MITRE ATLAS and OWASP Generative AI risks. The practical controls these frameworks recommend include real-time prompt filtering, strict access management and continuous monitoring of model behaviour. For anyone using AI tools that touch personal or financial data, understanding which of these frameworks your provider follows is now a reasonable question to ask. Explore more AI tools and tips in our Consumer AI section.



