--- title: "LiquidAI's ShieldFlow Scrubs Sensitive Data Before It Reaches the Cloud" subtitle: "Liquid AI's ShieldFlow scrubs sensitive data on-device before any prompt reaches the cloud, with no GPU required and a debut at Microsoft Build 2026" url: "https://alphasignal.ai/news/liquidai-s-shieldflow-scrubs-sensitive-data-before-it-reaches-the-cloud" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Liquid AI" company_url: "https://alphasignal.ai/company/liquidai" published: "2026-06-04T15:16:14.000Z" updated: "2026-06-16T20:21:42.756Z" type: "News" topics: - "Security" - "Gpus" subtopics: - "Content Moderation" - "Edge Deployment" - "Quantization" upvotes: 342 read_time: "6 min" source_url: "https://www.liquid.ai/request-access-shieldflow" comments: 0 --- # LiquidAI's ShieldFlow Scrubs Sensitive Data Before It Reaches the Cloud > Liquid AI's ShieldFlow scrubs sensitive data on-device before any prompt reaches the cloud, with no GPU required and a debut at Microsoft Build 2026 Every time a user pastes a document into an AI assistant, there is a moment where sensitive data crosses a boundary it probably shouldn't. Names, social security numbers, financial records, health information , all of it travels to a remote server before the model ever sees the prompt. [Liquid ShieldFlow](https://www.liquid.ai/request-access-shieldflow) is Liquid AI's answer to that problem: a privacy layer that runs entirely on the user's device and scrubs sensitive data before any prompt leaves the machine. ## A guard at the gate, not the cloud The core idea is straightforward. ShieldFlow is an on-device privacy layer that redacts sensitive data and prevents prompt injection before any prompt leaves the device. It sits between the user's input and whatever AI system they're calling , whether that's a local model or a cloud API , and sanitizes the text in real time. What makes this technically interesting is what's powering it. ShieldFlow runs on a device-native Liquid Foundation Model (LFM) , Liquid AI's proprietary model architecture , rather than a rules-based regex engine or a cloud-side classifier. That means it can understand context, not just pattern-match on known PII formats. The tradeoff with most on-device AI is hardware requirements, but Liquid is explicit that ShieldFlow needs no GPU and is light on memory, making it viable on ordinary laptops and PCs. ## The architecture behind it Liquid Foundation Models leverage Liquid Neural Networks, a proprietary architecture rooted in dynamical systems and signal processing, to deliver frontier-grade intelligence at a fraction of the compute, on any hardware, on or off the cloud. The LFM2 generation that underpins ShieldFlow is a hybrid model combining short convolutions with grouped query attention , a design that trades some of the memory overhead of pure transformer architectures for significantly better CPU performance. LFM2 achieves three times faster training compared to its previous generation, and benefits from up to two times faster decode and prefill speed on CPU compared to Qwen3. That speed advantage on CPU is precisely what makes a real-time, no-GPU privacy layer feasible. Whether deployed on smartphones, laptops, or vehicles, LFM2 runs efficiently on CPU, GPU, and NPU hardware. ## What it actually does ShieldFlow's job is to intercept prompts before they exit the device. In practice, this means: - Detecting and redacting personally identifiable information (PII) such as names, addresses, phone numbers, financial data, and health records - Blocking prompt injection attacks , attempts by malicious content in a document or webpage to hijack an AI agent's behavior - Running the sanitization step locally so the raw data never travels over the network - Allowing enterprises to deploy customized LFMs tuned to their own redaction policies Through Foundry Local, ShieldFlow runs efficiently on CPUs on every Windows device including AI PCs, and enterprises can pull a customized Liquid Foundation Model tuned to their own policies and roll them out across their Windows fleet through a single managed runtime. ## The Microsoft and AMD connection The timing of this launch was deliberate. ShieldFlow was featured at Microsoft Build 2026 as part of the [Foundry Local](https://devblogs.microsoft.com/foundry/accelerate-edge-ai-development-with-foundry-local/) ecosystem , Microsoft's cross-platform local AI runtime that lets developers ship AI features with no cloud dependency and no per-token costs. It also ran live on AMD laptops at Computex Taipei, demonstrating real-world performance on consumer hardware. At Microsoft Build 2026, Microsoft announced updates across Foundry Local and Foundry Local on Azure Local that help developers build once and run AI closer to where data is created and decisions are made, expanding platform support, improving control over inference and acceleration, and simplifying deployment across disconnected, regulated, and sovereign environments. ShieldFlow fits directly into that story as the privacy enforcement layer for enterprise AI PC deployments. ## Why this matters now The broader context here is a structural shift in where AI inference happens. Shifting large generative models from distant clouds to lean, on-device LLMs unlocks millisecond latency, on-device resilience, and data-sovereign privacy. But that shift creates a new problem: agentic AI systems that read documents, browse the web, and process user data need a trust boundary. ShieldFlow is designed to be that boundary. The prompt injection angle is particularly relevant for anyone building agents. As AI systems gain the ability to act autonomously , reading emails, summarizing contracts, filling out forms , the attack surface grows. A malicious instruction embedded in a document can redirect an agent's behavior in ways the user never intended. Catching that on-device, before the prompt reaches the model, is a fundamentally different security posture than relying on the model itself to resist manipulation. Instead of relying on cloud servers, these models are built to work locally, allowing AI systems to respond faster and keep user data on the device itself , an approach increasingly attractive as companies look for alternatives to always-online AI systems that raise privacy, latency, and long-term cost concerns. ## Who should pay attention ShieldFlow is most immediately relevant to: - **Enterprise IT teams** deploying AI assistants to employees who handle regulated data (healthcare, finance, legal) - **Developers building agentic pipelines** that ingest untrusted documents or web content - **ISVs building on Windows** who want a managed, policy-configurable privacy layer without building their own - **Organizations in regulated industries** subject to GDPR, HIPAA, or similar frameworks where data minimization is a design requirement The limitation worth noting is that ShieldFlow is still in early access. There is no public SDK, no published benchmark for redaction accuracy, and no pricing information available yet. The enterprise customization story , where companies tune the LFM to their own policies , is compelling but unproven at scale. ## How to get access ShieldFlow is currently in early access. Interested teams can [request access directly on Liquid AI's website](https://www.liquid.ai/request-access-shieldflow). For developers already building on Microsoft's stack, ShieldFlow is accessible through [Foundry Local](https://devblogs.microsoft.com/foundry/accelerate-edge-ai-development-with-foundry-local/), which is generally available on Windows, macOS on Apple Silicon, and Linux x64. Liquid AI's broader model ecosystem, including the LFM2 family, is available on [Hugging Face](https://huggingface.co/LiquidAI) for developers who want to explore the underlying architecture. The on-device AI wave is producing a lot of efficiency claims. What ShieldFlow adds is a specific, enterprise-grade use case: making it safe to run AI on sensitive data without trusting the cloud to handle it responsibly. Whether the redaction quality holds up under real-world enterprise workloads is the question that early access will answer. ## Sources - [X post by @liquidai](https://x.com/liquidai/status/2062554024436482301) - [Request access ShieldFlow — Liquid AI](https://www.liquid.ai/request-access-shieldflow) - [Accelerate Edge AI Development with Foundry Local — Microsoft Foundry Blog](https://devblogs.microsoft.com/foundry/accelerate-edge-ai-development-with-foundry-local/) - [Introducing LFM2: The Fastest On-Device Foundation Models on the Market — Liquid AI](https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models) - [Liquid AI LFM2.5 is Driving the Shift to Privacy-First AI — Times of AI](https://www.timesofai.com/news/end-of-the-upload-privacy-first-ai-liquid-ai/) - [Liquid Foundation Models — Liquid AI](https://www.liquid.ai/models)