.png)
As GenAI captures the world’s imagination, it’s also capturing the attention of attackers exploiting its flaws. Even the industry giants are not immune to it. In January 2025, researchers uncovered EchoLeak, a zero click vulnerability in Microsoft 365 Copilot that enabled attackers to exfiltrate sensitive data without any user interaction. Just a month before that, a prompt injection flaw in DeepSeek, was exploited to execute malicious JavaScript, leading to session hijacking and account takeovers. The number of reported AI related incidents increased by 56.4% in 2024, according to the Stanford 2025 AI Index Report. From phishing and model manipulation to guardrail bypasses and disinformation botnets, the threat landscape is evolving fast. As GenAI becomes embedded in mission critical systems across enterprise, financial services, healthcare etc, security can no longer be an afterthought, it must be foundational.
GenAI security is fundamentally different and far more complex than traditional security. The combination of dynamic model architectures, multimodal inputs, and autonomous agent workflows creates an ever expanding and unpredictable attack surface. Repello is tackling this challenge head on with a proactive, enterprise grade security platform purpose built for GenAI systems. Drawing on insights from red teaming thousands of AI systems, Repello has developed “ARTEMIS - the Autonomous Red Teaming Engine for Mapping, Identification, and Scanning”. Unlike traditional scanning tools, ARTEMIS brings deep business context to security testing, simulating realistic, domain specific attack scenarios powered by a proprietary threat intelligence corpus containing millions of attack vectors. With 15x the coverage of manual AI red teaming, <2% false positive rates post deployment, and vulnerability detection turnaround in under 6 hours (vs. 2+ weeks manually), ARTEMIS embeds directly into the development lifecycle enabling real time testing of inputs, prompts, and data across image, audio and text modalities in over 100 languages. The platform is now being extended to support AI agent security, marking a new frontier in GenAI risk management. Once vulnerabilities are identified, “Repello Guard” takes over offering a low latency, real time security layer that actively protects GenAI applications from prompt injections, data leaks, and exploitation. Together, ARTEMIS and Repello Guard represent a full stack, forward looking approach to GenAI security built not just to catch up with threats, but to stay ahead of them.
Since its founding a year ago, Repello has consistently demonstrated an unmatched ability to uncover critical vulnerabilities in GenAI systems across modalities. From bypassing advanced safety filters in Gemini 2.5 Pro, Claude 3.5 and ChatGPT using novel jailbreak prompts, to injecting malicious instructions via emojis and even ambient background noise, their research highlights how easily today's models can be manipulated in ways that evade human detection. They've exposed the fragility of popular diffusion models like Ghibli Dreams to adversarial attacks, and shown how safeguards like Meta’s “Prompt Guard” can be broken with precision crafted inputs.
Aryaman and Naman, co-founders of Repello AI, have known each other for more than five years and this is their second entrepreneurial initiative. Bug bounty hunters since college days, their passion for cybersecurity led them to start RepelloAI. We have been impressed by the drive and clarity with which they are building the company. Their community first approach is evident in their open source tools like Agent-Wiz, which maps complex AI agent interactions for threat modeling, and Whistleblower, which reveals system prompt vulnerabilities in LLM applications.
GenAI is transforming industries, but its long term impact hinges on how well we secure it. In a space where models evolve fast and threats evolve faster, Repello AI’s proactive & contextual platform and deep focus on adversarial R&D will give them a clear edge. Beyond category creation, Repello AI is architecting the security DNA for the future AI, where today's vulnerabilities could cascade into systemic risks. Their vision is to create an adaptive, self-evolving security mesh that can anticipate and neutralize threats in AI systems we haven't even imagined yet. We are excited to have partnered with them so early in their journey!
At pi, our thesis is to back deeptech companies building global businesses. While AI has been democratized to an extent, we believe there are still areas like AI security, privacy preserving techniques, specialized domains, to name a few, where there is an opportunity to build meaningful IP and defensible advantages. Repello aligns well with this thesis.