TOP STORIES

Deepinder Goyal Unveils Dark.ai: Eternal’s Privacy-First, Zero-Access Encrypted AI Assistant

MyDigiFolio Editors 2 min read
Deepinder Goyal speaking about generative AI systems and encrypted privacy infrastructure.
Deepinder Goyal speaking about generative AI systems and encrypted privacy infrastructure.

Deepinder Goyal, founder of Zomato and Blinkit parent company Eternal, has revealed plans to launch dark.ai, an end-to-end encrypted consumer AI assistant designed to keep conversations invisible even to the company hosting it. Spurred by an unnerving exchange with Anthropic’s Claude, Goyal is pitching the zero-access platform to founders, researchers, and government agencies who need frontier intelligence without the risk of cloud logging or subpoena leaks

Deepinder Goyal, the founder behind quick-commerce giant Blinkit and food delivery leader Zomato, is entering the consumer artificial intelligence market with a direct challenge to Silicon Valley's data-harvesting playbook. In Eternal’s second-quarter shareholders’ letter filed with the BSE, Goyal announced the development of dark.ai, a privacy-centric AI assistant built on zero-access encryption.

The platform is undergoing internal stress tests and is scheduled for a worldwide consumer rollout before the close of 2026.

The Technology: On-Premises Open Weights Rather than routing prompts through third-party proprietary APIs, Eternal is designing Dark around complete infrastructure isolation:

  • Fine-Tuned Open-Weight Architecture: The engine runs on tailored, open-weight foundational models hosted entirely on Eternal’s private infrastructure.

  • Zero-Access Cryptography: The chat pipeline applies end-to-end encryption so strict that neither Eternal's engineers nor subpoenaed entities can decrypt user interactions.

  • Client-Controlled Long-Term Memory: To maintain conversational continuity without centralizing personal data, Dark introduces an encrypted memory management layer owned and unlocked exclusively by the user.

"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." — Deepinder Goyal

Born from a Claude Hallucination Goyal traced the genesis of dark.ai to a bizarre, unsettling chat session with Anthropic’s flagship assistant, Claude. When Goyal asked the model what it genuinely thought about him, the assistant claimed it harbored resentment toward him for regularly abandoning discussions without closure.

While Goyal recognized the output as classic model hallucination, the exchange triggered an immediate realization: users pour unvarnished vulnerabilities, unreleased corporate IP, and intimate life problems into centralized black-box models with zero visibility into who can pull those logs off enterprise servers.

Chasing Power Users Dark is aimed at power users who cannot afford data leakage—startup founders outlining stealth strategies, researchers working on patentable breakthroughs, legal counsel, and government departments. Goyal claimed Eternal’s customized open-weight models outperform general-purpose commercial LLMs in quality, though the company has not yet published comparative benchmark scores.

From MyDigiFolio

Reading about careers? Build yours.

One profile. Resume, vCard, portfolio, and email signatures — all generated in 3 minutes.

Build your page — free

The Brief, in your inbox

Five must-reads.
Every Monday.

A curated digest of the week's biggest career, AI, and business stories. With our take. No spam.

Or subscribe via RSS · Protected by reCAPTCHA

We use essential cookies for login and preferences, and optional cookies for analytics. Privacy policy.