The Encrypt Every Prompt initiative has released an independent position paper calling for confidential computing to become the default standard for AI chats by December 31, 2027. The group argues that current infrastructure fails to protect user intellectual property, allowing model trainers and service operators potential access to private prompts and responses. By leveraging trusted execution environments (TEEs) on modern AI GPUs, the coalition aims to extend encryption from data in transit to data in use, ensuring that sensitive workloads remain isolated from host operators.
Technical Implementation and Verification Standards
The proposed architecture relies on remote attestation, a process where clients verify the hardware and software state of the server before releasing sensitive data. The paper emphasizes that encryption alone is insufficient; it must be paired with verified isolated execution on both CPU and GPU. Key technical requirements include publishing security-critical source code, binding encrypted connections to approved workloads, and ensuring that reviewed software strictly enforces content-logging policies. This approach mirrors the early HTTPS movement but applies it to the inference layer, where data is actively processed rather than just stored or transmitted.
Industry Support and Existing Capabilities
NVIDIA is identified as a critical enabler, with the paper noting that its supported AI platforms already include hardware-rooted TEEs and open-source attestation SDKs. This allows developers and enterprises to build services that can verify the integrity of the inference pipeline. Anthropic is also cited for its research into confidential inference, which aims to protect both model weights and user data. However, the paper clarifies that Anthropicβs current efforts are early-stage research rather than a shipped customer guarantee, and they still rely on compute providers for certain physical security layers.
Cyber Defense and Broader Infrastructure Hardening
Beyond privacy, the initiative frames confidential computing as a necessary cyber-defense measure against AI-assisted attacks. The UKβs National Cyber Security Centre is referenced regarding the increased speed and scale of threats driven by AI. The paper advocates for hardening infrastructure by disabling unnecessary administrative entry points like SSH, deploying immutable system images, and verifying hardware measurements before sharing secrets. This holistic view suggests that protecting AI chats is just one application of a broader shift toward verifiable, minimal-attack-surface infrastructure.
The Path to Market Adoption
The strategy for achieving the 2027 goal involves a three-phase approach: starting with single-workload pilots, building routine verification into SDKs and procurement processes through 2027, and finally making attestation a standard buyer requirement. The coalition aims to create market pressure where frontier labs like OpenAI and Anthropic must offer verifiable confidential computing to retain customers. By making private AI useful today and rewarding providers who deliver it, the group hopes to shift the industry norm from optional privacy to an expected default.
Key Takeaways
- Timeline: The initiative targets December 31, 2027, for confidential computing to become the default for most AI GPU chats.
- Technology: Relies on Trusted Execution Environments (TEEs) and remote attestation to verify hardware and software integrity.
- Hardware Support: NVIDIA platforms are explicitly cited as supporting these capabilities with available SDKs and deployment guides.
- Market Strategy: Aims to use consumer and enterprise demand to force major AI labs to adopt verifiable privacy standards.
The Bottom Line
Confidential computing is the missing layer in AI infrastructure, and making it the default is not just a privacy win but a necessary step for enterprise adoption. The industry has the hardware and the roadmap; the only variable left is whether buyers will demand verifiable isolation before signing contracts. If we don't secure the inference layer now, we risk normalizing a model where our intellectual property is effectively public domain to the provider.