“User-Centric AI: Building a Verifiable and Secure Future”

A reference for participants of the keynote by Illia Polosukhin, co-founder of NEAR Protocol, on October 1, 2026.

1. Main Theme

The central theme of the keynote, “User-Centric AI: Building a Verifiable and Secure Future,” is the critical need to shift from provider-optimized AI to a user-centric model where security, privacy, and verifiability are foundational. As AI becomes deeply integrated into our business and personal lives, the current “black box” approach from model providers introduces unacceptable risks, from data breaches and malicious code injection to a fundamental lack of trust. The core message is that we must build an AI infrastructure that empowers users and businesses to verify exactly which model is running, ensure data is end-to-end encrypted, and trust that the outcomes are not tampered with. This is achieved by combining confidential computing hardware with decentralized trust mechanisms.

2. Subthemes and Chapters

The keynote was structured around identifying the core problems with the current AI paradigm and presenting a layered, verifiable solution.

Chapter 1: The Trust Gap in Modern AI

The speaker began by highlighting the growing gap between what users expect from AI and what providers deliver. While AI adoption is seeing explosive growth (a “15x growth in inference”), our reliance on opaque, provider-controlled models creates significant security vulnerabilities.

• Problem 1: The “Super-Heist” Risk of Open Models: We don’t fully know the training data or methods of many models. This creates opportunities for “super-heists”—subtle, malicious behaviors embedded in a model that only trigger in specific, high-stakes contexts (e.g., designing an update for a nuclear power plant).
• Problem 2: Malicious Injections via Routers: When you use an AI service, your prompt is routed through various internal systems. The speaker explained that a provider can insert malicious bot commands into the model’s output. Your agent, trusting the output, will then execute the command, potentially compromising your entire system. The data is not safe, and the results can be tampered with.

Chapter 2: The Solution: Verifiable, User-Centric AI

To solve these issues, the keynote proposed a new infrastructure for User-Centric AI. The goal is not to limit what AI can do, but to make it globally accessible and fundamentally trustworthy. This requires a full-stack approach, but the talk focused on verifiable inference.

• Pillar 1: Confidential Inference: The core of the solution is making inference verifiable and secure. This is achieved by combining confidential computing hardware (like NVIDIA’s Confidential Computing mode or Intel SGX) with blockchain components.
  • How it Works: User data is encrypted on the client side. It can only be decrypted inside a secure enclave on a GPU. After processing, the output is re-encrypted and sent back with a “cryptographic attestation”—a verifiable proof that a specific, untampered model ran on your specific data within a secure environment. This prevents man-in-the-middle attacks, model-swapping, and result tampering.
• Pillar 2: The Decentralized Confidential Cloud: To make this technology globally available and scalable, the speaker introduced the concept of a decentralized network of data centers all running confidential computing hardware. This creates a “meta-cloud” where workloads are routed to the nearest available secure node. Because the hardware itself prevents the data center operator from seeing the data or the model, it creates a trustless, global infrastructure.

Chapter 3: Securing the Agent Layer

The final part of the keynote addressed the agent layer, which is becoming increasingly autonomous and powerful. As agents gain access to our email, calendars, and internal systems, the risk of exploits grows exponentially.

• The Problem: Every external input (an email, a calendar invite) is a potential vector for attack. Malicious actors can craft inputs to trick an agent into executing harmful actions.
• The Solution: A Verifiable General Purpose Log: By running agents within the confidential, verifiable infrastructure, all actions they take can be recorded in a tamper-proof “general purpose log.” This provides corporations and governments an auditable trail of every AI-driven action, allowing them to scan for malicious behavior, analyze agent performance, and ensure compliance without compromising the privacy of the underlying data.

3. Inspiring Quotes

• “Right now, when models are being provided to you, they are optimized for the business that is providing them… We need to ensure that the user and the business can verify which model ran.”
• “The problem is that with open models, you don’t fully know what the training data was, how exactly it was trained. Because of this, it’s actually important to have cryptographic guarantees on how you run your model.”
• “This enables truly end-to-end encryption, where you can encrypt it on your side… and it will be only decryptable inside a specific GPU.”
• “The capabilities are there, the models are actually very capable… but we’re not allowing them access to our things because somebody can send you an email with a malicious insert.”
• “We’re trying to push for AI that is built with privacy and security from the ground up, and that enables something that I believe is a public good.”

4. Key Action Points

1. Audit Your Current AI Supply Chain: Identify all third-party AI models and services your organization uses. Question what guarantees you have regarding data privacy, model integrity, and result verifiability. Ask your providers what happens to your data and which models actually run your prompts.
2. Explore Confidential Computing: Begin investigating confidential computing technologies (e.g., NVIDIA’s confidential GPUs, Intel SGX) for your most critical AI workloads. Start a pilot project to understand how end-to-end encrypted inference could protect your sensitive financial, legal, or proprietary data.
3. Demand Verifiability for AI Agents: As you deploy AI agents, insist on solutions that provide a verifiable, tamper-proof log of their actions. Prioritize platforms that offer cryptographic attestation to ensure the agent’s behavior is auditable and secure from external manipulation.