How co/core's two orthogonal provider guarantees work: Secure Mode (hardware attestation) and the Confidential tier (operator-blind inference).

NVIDIA chips now encrypt AI workloads with Confidential Computing, protecting data during inference using memory-encrypted confidential virtual machines.

Cohere plans on making the full serving stack used in Model Vault open source so that independent auditors will be able to validate that it doesn't log, export, or leak data.

OpenAI announced it has identified and blocked a coordinated effort to illicitly extract protected inference results from its AI models. Part of this core activity was linked to Moonshot AI, the Chinese startup behind the chatbot Kimi. The unauthorized activity began in early July and surged to 16,000 requests from over 4,000 users in just two days. OpenAI stated that it ultimately identified activities related to over 15,000 users and completely blocked them by July 28. Experts point out that this is a technique known as adversarial distillation, where a company uses another's model outputs and inferences to train its own. Concerns are rising that such extraction of inferences poses the risk of competitors mimicking advanced capabilities without massive investments or safety measures. There was no encryption breach or unauthorized database access; instead, hidden inference results were revealed through dialogue manipulation. Among US AI companies, anxiety is growing that rivals might misuse models for cheap and rapid technological development. OpenAI shared these findings with other developers through the Frontier Model Forum and government information-sharing channels.

This study investigates the risks of exposing confidential chemical structures when machine learning models trained on these structures are made publicly available. We use membership inference attacks...

In this podcast episode James Wilson chats with Tinfoil co-founder Tanya Verma about how you can run a powerful LLM in the cloud without t [Read More]

The open vs. closed AI model debate misses the bigger issue. Confidential inference secures model weights and data during runtime.

Announcing a new collaborative research paper on Confidential Inference, a set of tools to improve the security of our model weights and of our users' data

Attestation-bound encrypted tensor transport for confidential ML inference over VSock/TCP. Binary framing, X25519+ChaCha20Poly1305 AEAD, 3-message attested handshake. - cyntrisec/confidential-ml-tr...

Google Cloud announces significant innovations in Confidential Computing to enhance data privacy for AI deployments. Confidential Computing uses hardware-based Trusted Execution Environments (TEEs) to cryptographically protect data while it's being processed. Global scale Confidential AI capabilities now allow AI inference and fine-tuning with enforceable privacy guarantees. New Confidential G4 VMs with NVIDIA RTX PRO 6000 Blackwell GPUs are available globally, offering accessible Confidential AI. These VMs, powered by AMD EPYC CPUs and AMD SEV, protect data during processing and encrypt data transfer between CPU and GPU. Open-source Prompt Encryption SDKs provide end-to-end cryptographic protection for AI prompts and responses. Google Cloud is also collaborating with Apple to extend Apple's Private Cloud Compute on its platform, leveraging Confidential Computing and Intel TDX. Intel TDX is coming soon to C4 machine series Confidential VMs, providing hardware-isolated Trust Domains. Live Migration on C3D-based Confidential VMs is now generally available, enabling maintenance without workload interruption. Confidential Space, designed for secure multi-party computation, now integrates with Intel Trust Authority for independent verification. Additionally, Confidential Space now supports NVIDIA Hopper GPUs for secure, multi-party AI and machine learning workloads. These advancements aim to make Confidential Computing a foundational layer for secure collaboration and private AI innovation.

Qualcomm has unveiled its AI200 and AI250 rack-scale AI inference solutions relying on data center-grade Hexagon NPUs with near-memory computing, micro-tile inferencing, and confidential computing support.

Google's shipping private inference product runs Gemini inside hardware enclaves on custom chips, which is confidential computing rather than homomorphic encryption, and the company's actual homomorphic work is an unsupported research compiler.