Co-Founder and CEO, Adversa AI. Forbes Technology Council Member. Co-Lead, Agentic AI Security Workstream, Coalition for Secure AI (CoSAI). Co-Chair, IEEE Subcommittee for Cyber Security for Next Generation Connectivity Systems.
Alex Polyakov is the co-founder and CEO of Adversa AI, a pioneering startup in AI red teaming and agentic AI security. With over 20 years in cybersecurity, Alex began as a researcher uncovering over 300 zero-day vulnerabilities and evolved into a serial entrepreneur and trusted AI evangelist. As a Forbes Technology Council member, co-lead of the Agentic AI Security workstream at the Coalition for Secure AI (CoSAI), and co-chair of the IEEE subcommittee for "Cyber Security for Next Generation Connectivity Systems," he shapes the future of secure AI and emerging technologies.
Alex has spoken at over 100 global conferences, authored two books, and developed the first massive open online course on AI security. His team at Adversa AI has exposed critical vulnerabilities in models like xAI's Grok, ChatGPT, Claude, and DeepSeek, reinforcing the need for robust AI defenses. Through his work, Alex is dedicated to building trust in AI by protecting it from cyber threats, privacy issues, and safety incidents.
Polyakov opened with a live demonstration mindset: what if your AI product could be hacked in seconds? Over two decades at the intersection of cybersecurity and machine learning, he and his team at Adversa AI have done exactly that, exposing critical vulnerabilities in Grok, ChatGPT, Claude, and DeepSeek. Red teaming, the practice of simulating adversarial attacks against your own systems, is the only reliable way to discover these weaknesses before a bad actor does.
The masterclass walked through the architecture of an AI red-teaming program for early-stage and growth-stage AI companies. Polyakov explained the main attack surfaces that matter most: prompt injection, jailbreaks, privacy leakage, and model inversion. He referenced the OWASP Top 10 for LLM Applications, MITRE ATLAS, and the Coalition for Secure AI guidelines as the current baseline standards, then showed through real case studies, including the Amazon Q incident and the DPD chatbot failure, where companies went wrong and what a properly instrumented system would have caught first.
He closed with a framework for operationalizing AI red teaming inside a startup: what to test, how to structure continuous testing rather than one-off audits, and how to communicate findings to boards and regulators. The session's core message: AI safety and AI security are not separate disciplines; a model that can be jailbroken is a model that is unsafe, and every team deploying LLMs must treat adversarial robustness as a first-class engineering requirement.
Connect with AI safety researchers, policymakers, and practitioners shaping the future of responsible AI.