An independent position paper
Encrypt Every Prompt.
Our goal: protect AI chats with confidential computing on most AI GPUs by the end of 2027.
Buy compute. Keep your ideas.
A prompt is someone's work.
A discovery. A business plan. A private conversation. A line of code that took years of experience to write.
We are putting more of our lives into AI. The next generation of models will help us think, build, diagnose, negotiate and invent. The value in those conversations belongs to the people having them.
Paying for an answer should buy you an answer. It should never quietly give a model trainer a free copy of your ideas, your customers' secrets or your company's intellectual property.
Access to intelligence should come with ownership of your own thinking.
Training on someone's data should be a separate, informed agreement: permission, clear terms and compensation. Our private work should stay outside the training pipeline unless we choose to put it there.
We can build that boundary into the machines. The technology is here. Now we need to make it ordinary.
Encrypt the work.Verify the machine.
Encryption protects data on the wire and on disk. Confidential computing extends protection into the place where AI actually runs: the CPU and GPU.
A trusted execution environment, or TEE, isolates sensitive execution from the host operator. Remote attestation lets a client check evidence about the hardware and approved software before it releases sensitive data.
- Protect the whole route.Keep prompts, responses, caches and transfers within a protected path, from the client to attested CPU and GPU execution and back.
Protect the whole route.
Keep prompts, responses, caches and transfers within a protected path, from the client to attested CPU and GPU execution and back.
- Make the software inspectable.Publish the security-critical source, build measurements and verification tools. Bind the encrypted connection to the approved workload.
Make the software inspectable.
Publish the security-critical source, build measurements and verification tools. Bind the encrypted connection to the approved workload.
- Keep content under the user's control.Reviewed software must enforce the content-logging and reuse policy. An enclave alone can still run code that copies your data.
Keep content under the user's control.
Reviewed software must enforce the content-logging and reuse policy. An enclave alone can still run code that copies your data.
- Make verification a condition of service.Verify the actual route. Stop when its evidence fails. Give customers explicit control over every fallback and external tool.
Make verification a condition of service.
Verify the actual route. Stop when its evidence fails. Give customers explicit control over every fallback and external tool.
The boundary matters: a confidential gateway is only one part of the route. Hardware and firmware flaws, compromised clients, untrusted software and external tools still matter. Review each service's threat model, retention terms and exact model route. Read a detailed threat model.
Harden our systems for AI-powered attacks.
Confidential computing should become a cyber-defense priority far beyond AI.
AI-assisted attacks are raising the stakes. The UK's National Cyber Security Centre expects AI to increase the speed and scale of attacks. We should use this moment to make hardened, verifiable infrastructure easier for everyone to adopt.
- Close administrative entry points.Disable SSH and production debug shells. Keep only the services and processes the workload needs.
Close administrative entry points.
Disable SSH and production debug shells. Keep only the services and processes the workload needs.
- Boot a known image.Deploy reviewed, versioned images with immutable system partitions. Rebuild and redeploy updates rather than modifying live machines.
Boot a known image.
Deploy reviewed, versioned images with immutable system partitions. Rebuild and redeploy updates rather than modifying live machines.
- Verify before sharing secrets.Check hardware, software measurements and security settings against an approved policy before releasing keys or data.
Verify before sharing secrets.
Check hardware, software measurements and security settings against an approved policy before releasing keys or data.
- Protect the running workload.Use hardware-backed isolation and protected memory to limit access from compromised hosts and infrastructure operators.
Protect the running workload.
Use hardware-backed isolation and protected memory to limit access from compromised hosts and infrastructure operators.
AI can help drive adoption of this approach across computing. Better awareness, open tooling and practical deployment guides can help more teams reduce their attack surface and verify what is running. Protecting AI chats is one important use; stronger infrastructure is the wider benefit.
Confidential computing provides isolation and attestation; shell restrictions, minimal services and immutable images are additional controls to select and enforce. Google's production Confidential Space design shows how they can work together. Secure application code, patching, identity controls and scoped agent permissions remain essential.
NVIDIA supports it. You can build with it.
NVIDIA builds confidential-computing capabilities into supported AI platforms. Its plain-language explanation describes how hardware-rooted trusted execution environments protect models and data during processing. Its technical overview explains the deployment options and what each one protects.
The implementation path is public. Developers, enterprises and independent operators can follow NVIDIA's deployment guides and use its open-source attestation SDK to build and verify their own services. Or start with a confidential AI service available today.
Building your own requires a supported hardware and software combination, enforced security policies and the applicable production license. The reviewed workload and encrypted client connection determine whether the protection also excludes the model or service operator.
The major clouds support it, too.
The hyperscalers already offer confidential-computing infrastructure. AWS provides the Nitro System and Nitro Enclaves for isolated, attestable workloads. Google Cloud offers Confidential VMs (CVMs). Microsoft Azure offers confidential VMs, containers and supported confidential GPU configurations. Teams can build on these services today.
These are building blocks. Protecting an AI chat depends on the reviewed workload, attestation policy and encrypted client connection. GPU-backed inference also requires a supported confidential GPU configuration.
Use it today.
Private AI is already available. These are places to start, across enterprise buying, developer APIs and everyday chat. Select the confidential offering and check its verification requirements.
NY Token Exchange
Enterprise access to AI inference. Start with the confidential API and choose suppliers against your security requirements.
TrustedRouter
One API across models and providers. Choose a confidential route and check the gateway and upstream evidence.
Maple
Private AI for everyday conversations and work, with apps and developer tools built around confidential computing.
Privatemode
Confidential AI chat and APIs with client-side encryption, remote attestation and verification against published source.
Tinfoil
Private chat and inference on confidential hardware, with client verification and published workload measurements.
NEAR AI
AI chat and inference with confidential-computing options. Choose the Private TEE route and inspect its attestation support.
And growing. The goal is an open ecosystem.
Inclusion is an invitation to evaluate a service, rather than a certification or a claim that its team endorses this position. Offerings and guarantees vary by route.
Open models can move the whole market.
Start with open-weight models. Make encrypted AI chats, backed by verified confidential computing, an everyday experience.
As more people and companies choose those services, their spending creates demand for the same protection everywhere. Every workload moved and every contract that requires attestation gives the next provider a reason to offer it.
Our aim is to make that demand strong enough that OpenAI, Anthropic and other frontier labs have to offer verifiable confidential computing to win those customers. People should be able to choose the model they want while keeping their conversations private.
First an option. Then an expectation. Eventually, the default.
That is how we intend to change the market: make private AI useful today, reward the providers that deliver it, and make privacy a condition of winning tomorrow's business.
Protect customers' secrets, too.
Anthropic already makes the case. Its research on confidential inference describes work to protect both model weights and user data, using measured software and controlled release of decryption keys.
Our ask to the labs: extend that work into a customer-verifiable service. Give customers attestation, a connection bound to approved code, and clearly stated protections for prompts and responses across the whole route. Make the protection available to customers. Then make it the default.
Anthropic presents this as early research, rather than a shipped customer guarantee. Its interim design still relies on compute providers for physical and hypervisor security; detailed API-server protection is outside the article's scope. Closing those gaps is part of the work.
Make confidential computing the default for AI chats.
By the end of 2027, we want AI chats protected by verifiable confidential computing on most AI GPUs. The goal is to protect people's prompts, responses and context while they are processed.
Move one workload.
Run a real task on a verified confidential route. Compare cost, quality, latency and evidence.
Make it routine.
Put verification into SDKs, route selection and procurement. Publish the evidence and the gaps.
Make it the norm.
Buyers ask for attestation. Operators protect AI chats during inference. Model builders compete on intelligence.
This is a goal, rather than a claim about adoption today. Counting progress requires a published definition of protected AI chats, an inventory of GPUs serving those chats and evidence that the conversations stay within verified confidential execution.
A better bargain for everyone
More intelligence.Ownership of your ideas.
Builders: make private AI easy. GPU operators: make it available. Buyers: make it a requirement. Together, we can change what people expect from AI.