Eternal founder Deepinder Goyal said the company is building Dark.ai around encrypted conversations and user-controlled memory after an exchange with Claude sharpened his concerns about AI privacy.
In an Oct. 21 shareholder letter for Q2 FY27, Goyal said Claude resented him for leaving conversations unfinished. He said he was unsure how to interpret the response and thought it could have been confabulation.
The planned service would pair end-to-end encryption with a memory feature users control, which Goyal said is intended to carry context between sessions. It would also run fine-tuned open-weight models on Eternal’s infrastructure.
Goyal said Eternal would be unable to read users’ chats even if compelled. “We cannot read your conversations, even if we are forced,” he wrote. The letter did not explain the encryption architecture or cite an independent security audit.
Goyal’s concerns extend beyond whether chats are used to train AI models. He cited possible data storage, legal holds, human audits, abuse reviews and outside logging as other ways information could be handled. He argued that opting out of training would not address every access concern, and said uncertainty could discourage people from sharing sensitive work or asking personal questions. He did not tie those examples to a particular provider.
An earlier EconoTimes report on Dark.ai’s business status said Chief Financial Officer Akshant Goyal told analysts on Oct. 21 that the project had no revenue or settled business model.
Deepinder Goyal named founders, researchers and scientists, as well as enterprises and governments, as intended users. He said Eternal is testing Dark.ai internally and plans to share model benchmarks soon. The company hopes to launch globally before the end of 2026, but Goyal gave no specific release date.
