The math is simple -> more rewards = more inference + you keep your NEAR.

Most AI privacy comes down to a policy you have to trust. With Leo on @near_ai, your prompt is encrypted in Brave and only opens inside an enclave that signs a proof of where it ran and which model answered. We walk through how it works and how to set it up: youtu.be/Shxc59Rkg68

Builders! ⚠️ Deprecation announcement ⚠️ The Graph will start winding down their integration with NEAR blockchain from 16th Oct onwards if you are using their service in any way, you should consider migration to an alternative: • Goldsky indexer: they have a guide for the migration from The Graph - docs.goldsky.com/subgraphs/m…. • Neardata API - docs.fastnear.com/neardata

Shipping a mobile app on NEAR AI? Your API key is extractable the moment someone downloads it. @joespano_ did a walkthrough on how to keep it off the binary and your prompts out of any middleman. youtu.be/bzTljJk_dNs

If you see a builder, nominate them to create a profile with @NearBuilders Its a great way to connect and get discovered on NEAR

You can remove the privacy concern by using private inference with @near_ai

Leveraging AI tools also takes consideration about talent development. Joe got a chance to talk about this with TalentObserver.

Want to learn how to build with private AI? It's easy to get started so we built a bank statement analyzer on NEAR AI cloud and recorded the whole thing. Upload a statement, model processes it inside a TEE, get back categorized transactions plus a proof panel. Whole run cost two cents. Watch how easy it is to get started. youtu.be/jQ33aPEw_mU

Every AI app says it protects your data. Here's what that looks like in actual code. Four lanes, ranked by privacy. An audit record written before the route commits, not after. A ratchet so a conversation can move to a more private lane but never back. Plus the log chain running on the phone. youtu.be/m3L_4mqZuxA

In case you missed them, three videos on the same problem from different angles: how do you build an AI product without handing your users' data to whoever runs the model. Read in this order they build on each other. 1. Trusted Execution Environments, explained We encrypt data at rest and in transit. Then it reaches the model and has to be readable, because nothing can process ciphertext. That's the hole. A TEE seals the processor so the operator of the machine is locked out along with everyone else, and it hands you an attestation report you can verify against the silicon vendor. You're checking a promise rather than accepting one. youtu.be/QXCnL_7MihA 2. Switching to private inference The migration is a base URL change. NEAR AI's cloud API uses the same setup as OpenAI, so your calls start hitting a TEE-hosted model without a rewrite. The video also covers scope, which matters more than the switch itself. That guarantee applies to the model on that path, not your app, not your company. If you log the request before it leaves, the enclave can't help you. youtu.be/2fRt-W_w2L0 3. The safety layer that runs first This one argues against the previous two in a useful way. Before you call any model, private or otherwise, run an emergency check locally, then ask whether you need a model at all. Someone asking for their latest blood pressure reading is asking for data already sitting on their phone. That's a lookup. Adding an outside party to it, even a sealed one, makes it less private rather than more. youtu.be/Gx2i2PUZXHk Questions welcome.

The latest incident with user data and AI training continues to highlight the importance of privacy being built into the AI tools from the start.

NEP-641 has been just approved. "Offchain Authorizations for Smart Contracts" This standardizes the flow for signing and verifying offchain messages for smart contracts that power Passkey, EVM, multisig, and other wallet contracts. The first version of this standard was live and tested on Trezu, and now it is a standard for the whole NEAR ecosystem.

Privacy includes being able to recognize when something is insecure.

Ironclaw v1 is out and the benchmarks are impressive. But what is benchmarking actually telling us? Benchmarking is the difference between "trust me bro" and show me the proof. Every AI harness says theirs is the best, benchmarking is how you separate marketing claims from reality. Think of it like car ratings. You wouldn't buy a car based on a single test drive around the block. You want to know: does it brake on ice? Does the safety system actually work? What's the real MPG? Same with AI assistants. The best benchmarks test practical stuff: → Schedule a meeting from a messy email → Research a topic and write a summary → Book a flight on a real website → Find the right number across 90,000 pages of documents That last one? That's OfficeQA with 89,000 pages of US Treasury bulletins, 26 million numbers, much of it scanned PDFs. An AI that scores well there can handle your company's messy internal docs and certainly that pile of saved recipes you have too. IronClaw 1.0 just topped the leaderboard on three of these benchmarks (PinchBench, ClawBench, OfficeQA) using the same base AI model everyone else had. The difference is the architecture, think of it as a car that corners better because the chassis was redesigned, not just a bigger engine bolted on. Why should you care? Because without benchmarks, every AI is the "best" until you actually use it. With them, you know what you're getting before you trust it with real work. So go try it at ironclaw.com See the full breakdown: near.ai/blog/introducing-iro…