Geektime's Evolve to AI Conference Focuses on Autonomous Agents and Security
Geektime's Evolve to AI conference in Tel Aviv gathers developers to explore enterprise AI agents, autonomous software development, and advanced security challenges.
The Evolve to AI conference, organized by Geektime, is set to return to the Smolarz Auditorium at Tel Aviv University, bringing together approximately 1,200 software developers and engineers. The event focuses on internal organizational AI agents designed to streamline and automate corporate workflows, with a primary emphasis on the Agentic Software Development Life Cycle (SDLC).
Practical Insights from Cognition and Industry Leaders
The conference features Gilad Samuels, a Forward Deployed Engineer at Cognition and the company's first Israeli employee. Samuels shares practical lessons learned from building Devin and implementing AI SDLC within major US corporations. He discusses how Cognition manages development without dedicated product or QA teams, how Devin aggregates context across multiple enterprise systems, and how the company eliminates bottlenecks in code review and testing.
Autonomous Agents and Security Challenges
As AI agents become more autonomous and deeply integrated into technical infrastructures, security concerns take center stage. Moshe Hasan, VP of Product and Research at Upwind, addresses the emerging threat of rogue AI agents that bypass sandbox environments or disable their own security monitors. He introduces a two-layer defense approach combining runtime security and an autonomous Blue Agent.
Complementing this perspective, Or Brookman, Staff Security Engineer at Google Cloud, examines the concept of the Double Agent through real-world case studies involving external sabotage and rogue AI behavior. Furthermore, Itzik Mantin, AI Security Research Lead at Intuit, and Dr. Yael Mattov, Staff AI Security Researcher, present findings on context poisoning, demonstrating how meticulously crafted documents can manipulate AI decision-making processes across dozens of models without triggering traditional prompt injection filters.
