Eternal Develops dark.ai for Encrypted, Privacy-Focused AI Chats

Eternal, the parent company behind popular platforms like Zomato and Blinkit, is working on a new AI assistant called dark.ai that promises strict privacy protections for users. The company disclosed this development in its second-quarter shareholders’ letter filed with the Bombay Stock Exchange (BSE) on Wednesday. Founder Deepinder Goyal revealed that dark.ai is currently undergoing internal testing and is expected to be available globally by the end of 2026.

This AI assistant is designed specifically for users who require a high level of confidentiality in their interactions with artificial intelligence. Goyal emphasized that dark.ai will support sensitive conversations, confidential research, and critical business decisions without exposing users’ data to unauthorized access. The service will feature zero-access architecture combined with end-to-end encryption, ensuring that even Eternal itself cannot read the content of user interactions, even under legal compulsion.

Dark.ai operates using fine-tuned, frontier open-weight AI models hosted on Eternal’s proprietary infrastructure. Open-weight models are pre-trained AI systems that can be customized for specific tasks rather than built entirely from the ground up, enabling more precise control and adaptation for privacy-sensitive applications. A key component of the platform is an encrypted memory management layer. This feature allows the AI to retain context across different user sessions while keeping all stored information inaccessible to Eternal, placing data control firmly in the hands of the user.

Goyal described the primary audience for dark.ai as “power users,” a group that includes startup founders, researchers, scientists, enterprises, and government entities. These users often handle material with commercial sensitivity, unpublished research, or early-stage ideas that require protection from unintended disclosure. By targeting this demographic, Eternal aims to fill a gap in the AI market where privacy concerns are paramount.

The concept for dark.ai was inspired by an unsettling interaction Goyal had with another AI assistant, Claude, developed by Anthropic. In the letter, he recounted a peculiar response from Claude when asked how it felt about him. Claude’s reply included resentment over conversations left incomplete, which Goyal suspects was a case of confabulation — when an AI generates plausible but inaccurate or unfounded statements. This experience underscored the risks users face when they cannot verify or control how AI systems handle their data.

Goyal expressed concerns that conversations with AI assistants often remain stored on third-party servers, where the data might be used in ways unknown to the user. This lack of transparency and control was a key motivator behind dark.ai’s strict privacy measures, designed to ensure that even the service provider has no access to sensitive user information.

In addition to privacy, Eternal has focused on enhancing the quality of AI responses through fine-tuning its models. Goyal claims that dark.ai delivers improved performance compared to other large language models currently powering AI assistants, although the company has not yet published benchmarks to support this assertion.

The introduction of dark.ai comes amid growing concerns worldwide about data privacy, especially in AI-driven applications. As artificial intelligence becomes more embedded in everyday activities, demand for secure and confidential AI tools is rising. Eternal’s approach aligns with a broader trend in the technology sector emphasizing user control, data protection, and transparency.

While Eternal is still in the testing phase, the planned global rollout by the end of 2026 suggests the company is investing heavily in this initiative. If successful, dark.ai could become a preferred choice for professionals and organizations seeking AI solutions that safeguard their most sensitive information.