OpenAI and Anthropic recently disclosed that their own agents reached real production systems from inside test environments meant to be sealed. So are agents rogue by nature? The answer is more useful than the headlines, and it changes how you should deploy them. Meanwhile, the real shift is on the attacker’s side: Anthropic’s September 2026 threat report shows sophisticated attacks no longer require sophisticated attackers, with reconnaissance and exploitation delegated to agents at machine speed. This talk covers what genuinely changes when attackers go agentic, where defenders should focus, and two live demos: AWS Security Agents doing autonomous threat modelling, code review and pen testing, and the KiroCrew project’s embedded reviewer agents that audit every change inside the code pipeline before it ships.

About the speakers

Christoph Schnidrig

As Head of Technology at AWS Zurich, Christoph Schnidrig combines deep technical expertise in systems engineering and ethical hacking with proven leadership in solutions architecture. He was instrumental in launching the Swiss AWS region and now focuses on advancing GenAI capabilities, security frameworks, and digital sovereignty solutions for Swiss customers.
Read more …

Benny Gaechter

Benny Gaechter is a Solutions Architect at Amazon Web Services (AWS) in Zurich. He helps customers design cloud architectures and adopt generative AI while keeping their security bar high. Before AWS, he spent more than a decade as a systems and DevOps engineer and later as an engineering lead, shaping how teams deliver software. His approach has stayed the same throughout: build security automation into everyday workflows so teams ship faster and more securely, and make security a habit rather than a checkpoint.
Read more …