Most of the AI tools businesses have adopted so far handle low-risk work, such as drafting emails, summarizing calls and cleaning up slide decks.
The material companies would most like to hand over is the material they guard most closely.
That includes financial research before it’s published, legal files covered by attorney-client privilege and internal data on pricing, products and deals.
What holds that work back is usually not what the models can do. It’s what happens to the data once it leaves the building.
Sending a confidential document to an AI model means trusting the cloud provider, the chip operator and the AI company not to see it. Most of the time, that trust rests on a privacy policy.
Related: Bernie Sanders spots opportunity as 66% of Americans oppose AI data centers
The pressure is showing up in corporate budgets. In Cisco’s 2026 Data and Privacy Benchmark Study, 90% of the more than 5,200 IT and security professionals surveyed said their privacy programs have expanded because of AI. Cisco puts it plainly:
“AI ambition is outpacing readiness.”
The cost of getting it wrong is rising too. IBM’s 2026 Cost of a Data Breach Report puts the average breach at $4.99 million, a record. Attacks that pull sensitive data back out of AI models cost even more, at $6 million on average.
NEAR AI, the AI arm of the NEAR ecosystem, is building its business around replacing that policy with proof.
On October 15, the company announced that Leo, the AI assistant built into the Brave browser, now runs on private inference from NEAR AI Cloud.
Inference is the step where a model reads a prompt and produces an answer, and it’s where user data is most exposed.
A locked room the operator can’t enter
According to NEAR AI, Leo conversations on the GLM 5.3 Flash model run inside trusted execution environments, or TEEs.
A TEE is a walled-off section of a processor where code and data are shielded from the rest of the machine. NEAR AI says its TEEs are built on Intel TDX and NVIDIA Confidential Compute.
The company says the cloud host, the GPU operator and NEAR AI itself are “cryptographically locked out.” Leo encrypts each prompt in the browser with a protocol called Oblivious HTTP, using a key that exists only inside the enclave.
Brave’s relay strips the user’s IP address before passing the request on. NEAR AI says memory stays encrypted while the model runs, and every response is signed inside the enclave.
The bigger point is who does the checking. Before Leo receives the enclave’s key, Brave’s own open-source verifier checks the hardware against Intel’s signature and NVIDIA’s attestation service.
Attestation is a cryptographic proof of what hardware and software are running.
The verifier also rebuilds NEAR AI’s code from public source to confirm it matches what is deployed. As the post puts it: “No pass, no key.”
NEAR AI’s Trust Center puts it in business terms. It says privacy is “enforced by hardware, not policy,” and verifiable through attestation “checked against Intel and NVIDIA, not NEAR AI.”
Trending on TheStreet Roundtable:
Veteran economist says investors are watching the wrong bubble
Billionaire Ray Dalio sounds alarm on AI’s $4.4 trillion debt squeeze
Two big U.S.-listed Bitcoin miners plan major exit for AI
Brave’s reach as a distribution channel
Distribution may matter as much as the technology. Privacy tools usually struggle because users have to go find them. Inside a browser, private inference becomes a setting.
Brave currently serves over 130 million monthly active users. That figure shows the size of the potential audience.
Private inference applies only to Brave Leo users who choose GLM 5.3 Flash in the model settings.
The rollout came in stages. Brave began testing NEAR AI’s TEEs in November 2025 with DeepSeek V3.1 in Brave Nightly, its experimental build. The October release takes it to the full production browser.
Brendan Eich, co-founder and CEO of Brave, called the integration “a first for AI chats in browsers.” Illia Polosukhin, co-founder of NEAR, framed it as a question of ownership.
“Browsers are a primary computing interface that handles all of our personal information, so it’s critical that AI is introduced in a user-owned way,” Polosukhin said.
NEAR AI says Brave joins Venice AI, Abound and the Government of Bermuda among organizations running production workloads on NEAR AI Cloud.
That mix of consumer apps and a public-sector user hints at the same distribution model at work: privacy delivered through software people already use, rather than a standalone product.
Guarantees that stop at the model
For businesses weighing sensitive workloads, the fine print matters. The privacy guarantee in the
Brave announcement covers one model in one deployment. NEAR AI describes GLM 5.3 Flash as an open-weights model under the MIT license, meaning its underlying parameters are public.
NEAR AI’s Trust Center specifies that open-weight models hosted by NEAR AI run inside the TEEs. Third-party models are also offered, but NEAR AI says those are “proxied without revealing the customer’s identity and processed under the provider’s own data terms.”
In other words, the hardware guarantee applies only where NEAR AI controls the hardware.
That distinction will likely shape how companies buy. A guarantee tied to a named model, a named deployment and an attestation anyone can check gives them something to audit.
NEAR AI also says a hardware-attested deployment log confirms that the model a user picks is the one answering.
Publishing open-source verification and relying on chipmakers’ attestation services rather than its own word lets NEAR AI’s approach be tested from outside.
For businesses that have kept their most sensitive files away from AI, that offers a way to bring them in without taking anyone’s word for it.
Related: $13 trillion brokerage giant offers cheapest ways to trade crypto