INForum 2026

Oradores Convidados

• 

  Fred Oliveira 🔗

  Snyk

  Título

  A Whirlwind Tour of Open Problems in AI Security

  Resumo

  Over the last few years, we have seen large language models break many of the assumptions behind the trust mechanisms that make up the cybersecurity landscape: that humans would read code; that attacks wouldn't be intelligent; that the parts of a system could be named, versioned, and looked up in a database; that anything able to act on a machine would be a program you could scan; and that a human would be on the other end of every alarm. As a field and as an industry, it is on us to tackle these issues. This talk will be a tour of the open problems in AI Safety and AI Security. Some of these, I have field notes on. These are the problems my team and I are trying to solve at Snyk: from risk evaluations, to red-teaming applications, to securing the new agentic development lifecycle. Some of these problems, however, are open-ended, and require fresh perspectives. Maybe the perspectives of the people in the room. Expect to leave with perhaps more questions than answers. But right now, I believe that's the most useful thing I can offer.

  Bio

  Fred Oliveira works on AI risk and agentic security at Snyk, where he builds evaluation methodology and observability tooling for autonomous agent behavior. He previously built data systems for venture capital funds. He co-authored O'Reilly's Redis Cookbook and co-organizes LisbonAI. He is based in Coimbra.

• 

  Nuno Preguiça 🔗

  Faculdade de Ciências e Tecnologia -- Universidade NOVA Lisboa

  Título

  From Conflict Resolution to Application Correctness in Distributed Systems

  Resumo

  Distributed applications, running on multiple nodes, need to share data. When coordination is too costly or simply unavailable - from mobile computing to geo-replication setting and autonomous operation - systems resort to weak consistency, and concurrent updates must be reconciled after the fact. In this talk we explore conflict resolution, the family of approaches proposed for doing so. We show that resolving conflicts independently for each object is not enough to provide application correctness: convergence at the object level says nothing about the application invariants that span multiple objects and define what a correct application state actually is. We discuss proposals that address this gap and detail the No-Op framework, a generic approach for enforcing correctness without coordination while guaranteeing both convergence and invariant preservation.