Google Picks Mirror Security to Help Solve AI’s Plaintext Problem

Mirror Security has been selected for Google’s Gemini Startup Forum: Cybersecurity, putting the AI-security startup among 33 companies chosen to work more closely with some of Google’s core cloud, AI and cybersecurity teams.

The programme is designed for startups developing AI-enabled cybersecurity technologies. Over the course of the initiative, Mirror Security will receive access to technical expertise and product support from Google Cloud, Google DeepMind and cloud-security company Wiz, while exploring opportunities to build within the broader Gemini ecosystem.

For Mirror Security, the programme aligns closely with the problem it is trying to solve: how enterprises can use powerful AI systems without exposing sensitive information during computation. The company develops encrypted AI infrastructure using fully homomorphic encryption, or FHE, which is designed to allow models to process information while that underlying data remains encrypted.

Its broader platform also focuses on securing AI agents, controlling what tools they can access, enforcing runtime policies and creating audit trails around their actions. Mirror Security argues that conventional encryption protects information while it is stored or moving across networks, but sensitive data can become exposed once an AI model needs to process it — the gap its technology is built to address.

Founded in 2024, Mirror Security emerged from University College Dublin’s NovaUCD deeptech ecosystem. The company was cofounded by Pankaj Thapa and Dr. Aditya Narayana and is backed through Sure Valley Ventures’ third fund.

The Google programme gives Mirror Security more than visibility. Access to teams across Gemini, cloud infrastructure and cybersecurity could help the startup test how encrypted AI and agent-security products fit into larger enterprise environments. As companies deploy increasingly autonomous AI systems, protecting the data those systems access — and controlling what those agents are allowed to do — is becoming a cybersecurity problem of its own.