In this insightful discussion, Maxim Bar Kogan, CEO of Onyx Security, highlights the growing risks enterprises face as AI agents become more autonomous and capable. He explains that while enterprises are rapidly adopting AI agents like AutoGPT and Cloud Code to automate complex tasks, this exponential increase in agent actions also raises significant security concerns. Traditional security measures such as identity controls and endpoint protections are insufficient because these agents require broad permissions to function effectively, making it challenging to prevent harmful or unintended actions. Onyx Security addresses this by building specialized models and agents designed to oversee and control other AI agents, ensuring their actions remain legitimate without overwhelming human operators.
Maxim elaborates on Onyx’s approach, which involves training smaller, focused models that monitor agent behavior and decide when to escalate issues to more capable agents for deeper analysis. This layered oversight balances performance, cost, and reliability, allowing enterprises to scale AI adoption safely. He draws an analogy to blitz chess, where intuitive quick decisions are made most of the time, but critical moments receive more thoughtful attention. This strategy helps manage the complexity and unpredictability of autonomous AI systems while maintaining operational efficiency.
The conversation also touches on the unique strengths of the Israeli tech ecosystem, particularly its blend of cybersecurity expertise and AI research. Onyx’s team, with backgrounds in Israeli intelligence and AI, exemplifies this fusion, enabling them to tackle the complex challenge of AI oversight. Maxim emphasizes that understanding the day-to-day realities of security teams is crucial for building effective products, as these teams operate within structured organizations with specific workflows and pressures. This user-centric approach is a key differentiator for Onyx in the competitive AI security landscape.
Addressing the broader industry context, Maxim discusses the inevitability of more advanced AI models like Mythos and the challenges they pose, including the accelerated discovery of software vulnerabilities. He stresses the importance of foundational security measures tailored to AI’s unique risks and expresses cautious optimism about phased rollouts of advanced models by major labs. However, he believes that independent oversight companies like Onyx will remain essential because enterprises need trusted third parties to validate and govern AI behavior, especially as multiple AI vendors and open-source models proliferate.
Finally, Maxim shares his vision for the future, where security teams will increasingly be augmented or even run by AI agents themselves. Despite his strong belief in the eventual arrival of AGI, he underscores the importance of delivering practical, user-friendly solutions today for human operators while preparing for a future where AI agents become the primary users of security tools. This dual focus on present usability and future adaptability positions Onyx Security as a pioneering company in the evolving field of AI governance and enterprise security.
