Ep. 28 - The Attacker's Perspective on AI Security with Aryaman Behera

In episode 28, we talk to Aryaman Behera, Co-Founder and CEO @ Repello AI.

Repello AI is a startup that just raised a strategic seed round from investors like General Catalyst to build an automated AI red teaming and security platform. But Aryaman's story doesn't start in a VC pitch room. It starts in the trenches of offensive security. He's a former bug bounty hunter, the captain of India's top-ranked cybersecurity CTF team - InfoSecIITR. He's now taken that deep, attacker-centric mindset and is building tools to help enterprises find and fix the novel vulnerabilities that come with deploying LLMs at scale.

In this episode, Aryaman shares his unique journey from being a bug bounty hunter and the captain of India's top-ranked CTF team, InfoSec IITR, to becoming the CEO of an AI security startup. The discussion offers a deep dive into the attacker-centric mindset required to secure modern AI applications, which are fundamentally probabilistic and differ greatly from traditional deterministic software. Aryaman explains the technical details behind Repello's platform, which combines automated red teaming (Artemis) with adaptive guardrails (Argus) to create a continuous security feedback loop. The conversation explores the nuanced differences between AI safety and security, the critical role of threat modeling for agentic workflows, and the complex challenges of responsible disclosure for non-deterministic vulnerabilities.

Key Takeaways

• From Hacker to CEO: Aryaman discusses the transition from an attacker's mindset, focused on quick exploits, to a CEO's mindset, which requires patience and long-term relationship building with customers.

• A New Kind of Threat: AI applications introduce a new attack surface built on prompts, knowledge bases, and probabilistic models, which increases the blast radius of potential security breaches compared to traditional software.

• Automated Red Teaming and Defense: Repello’s platform consists of two core products: Artemis, an offensive AI red teaming platform that discovers failure modes , and

• Argus, a defensive guardrail system. The platforms create a continuous feedback loop where vulnerabilities found by Artemis are used to calibrate and create policies for Argus.

• Threat Modeling for AI Agents: For complex agentic systems, a black-box approach is often insufficient. Repello uses a gray-box method where a tool called AgentWiz helps customers generate a threat model based on the agent's workflow and capabilities, without needing access to the source code.

• The Challenge of Non-Deterministic Vulnerabilities: Unlike traditional software vulnerabilities which are deterministic, AI exploits are probabilistic. An attack like a system prompt leak only needs to succeed once to be effective, even if it fails nine out of ten times.

• The Future of Attacks is Multimodal: Aryaman predicts that as AI applications evolve, major new attack vectors will emerge from new interfaces like voice and image, as their larger latent space offers more opportunities for malicious embeddings.

We hope you tune in and, if you like the episode, please do subscribe!