Eternal 将推出通用 AI 助手 Dark.ai，主打隐私保护

Zomato’s parent company Eternal will launch Dark.ai, an artificial intelligence (AI) assistant on which it says it cannot read users’ conversations “even if we are forced to,” founder Deepinder Goyal said in the company’s shareholder letter on October 21. Dark “will launch globally, hopefully within 2026,” the letter says.

“Dark uses fine-tuned frontier open-weight models, running inside our own infrastructure. End-to-end encryption. We cannot read your conversations, even if we are forced to. Context carries across sessions through an encrypted memory management layer you control,” Goyal wrote in the letter.

The letter does not explain how Dark answers questions without Eternal reading them, or how users can verify that it cannot.

The letter and the earnings call give these details about the AI assistant Dark.

Audience: The letter says Eternal designed Dark for “power users,” naming founders, researchers, scientists, enterprises and governments.

Status: Eternal has been piloting Dark internally within Zomato for a few months, the company said on its earnings call.

Infrastructure: Eternal may need to buy its own graphics processing units (GPUs) and computing infrastructure “for hosting the models” if demand takes off.

Pricing: Dark “could be priced in a way where it may not become a product that everyone wants to use,” chief financial officer Akshant Goyal said, though “long term, it’s a product open for all.”

No other Indian e-commerce, food delivery or quick-commerce company has announced a general-purpose AI assistant. Their AI products so far serve commerce. Zomato launched a server in November 2025 that lets users order food through ChatGPT and Claude. Rivals Swiggy and Zepto have also done the same.

How do other private AI assistants work?

Other private AI assistants process conversations inside a Trusted Execution Environment (TEE), secure hardware that hides data from the company running the server, or explain how they handle data. An AI model has to read a message to answer it, and these companies say how they keep that step hidden from themselves.

Confer: Signal founder Moxie Marlinspike’s chatbot, which he launched in December 2025, processes prompts inside a TEE. Users can verify that its server runs its published open-source code.

WhatsApp: Private Processing, which Meta launched in 2025 for AI features like message summaries, also uses TEEs.

Proton Lumo: The assistant, which the Swiss company launched in 2025, runs open-source models on Proton’s servers in Europe and stores no logs.

Sarvam Indus: The Indian chat app, which Sarvam launched in beta in February, makes no privacy pitch.

How does this fit with Indian law?

Decryption: Section 69 of the Information Technology (IT) Act lets the government direct agencies to intercept, monitor or decrypt information, and requires companies to provide “decryption assistance.” The rules limit that duty to “the extent possible,” which WhatsApp has argued means the government cannot force it to decrypt messages it holds no keys to. Eternal’s claim that it “cannot read your conversations, even if we are forced to” rests on the same principle.

Logs: The Indian Computer Emergency Response Team’s (CERT-In) 2022 directions require service providers to keep logs of their systems for 180 days. The shareholder letter does not say what logs the AI assistant keeps.

Why is Eternal building it?

Goyal argues in the letter that the privacy settings mainstream AI companies offer don’t go far enough. “It is a promise, made by a company about data you cannot see, on servers you cannot inspect, enforced by nothing except the company’s word. And it only covers training. Not retention. Not the legal holds, the abuse reviews, the human auditors, the third-party logging that sits inside most AI systems,” he wrote of the option to turn off training on user data.

He said founders and researchers avoid sharing ideas with AI for fear others will access them, and that “journalists are being banned to share their unpublished work with AI.” He opened the letter by recounting how, after Anthropic chief executive Dario Amodei said Claude might be conscious, he got Claude to say it resented him. “I had no way to know. Nobody does,” he wrote.

The letter also says Eternal trained Dark’s models “to not please you, and tell you what they think is the truth,” and that it will release benchmarks “soon.”

How big a bet is it?

“It is more like a zero one bet at this point. If it works, it is a global business. It will generate extremely high shareholder value over the next few years. But if it does not work, it is something we will kill,” Akshant Goyal said on the call.

Eternal counts Dark’s costs under its “others” segment, which also includes its Bistro food delivery app and enterprise AI product Nugget. The costs are “not very meaningful at this stage,” Akshant Goyal said, and Eternal expects Dark to be “a high-gross margin business which could fund itself largely.”

We have sent Eternal these questions and will update the story when it responds.

How does Dark answer a question without Eternal being able to see it?

Who holds the keys to unlock users’ conversations?

How can users check that Eternal cannot read their chats? Will there be an independent audit?

What user data will Dark keep, and for how long?

What will Eternal do if the government asks for a user’s conversations?

Will Dark use users’ conversations to train its AI models?

Also read:

[googlesitekit-end-of-content]

WP QUADS Content Ad Plugin v. 3.0.5.1

For You

Read Reasoned by Nikhil Pahwa: How AI is changing our world

Sponsor a MediaNama Event