Issue Merging data and computation removes the separation classical security relied on

Summary

In a sidebar interview, Stanford assistant professor Sanmi Koyejo, cofounder of Virtue AI, explains why AI systems are harder to protect than conventional computing. The biggest difference, he says, is how much more flexible and contextual AI is, so tools perfected for traditional security work far less well on AI and even less on agentic systems. In classic computing, data and compute were kept apart, letting defenders distinguish attacks on data from attacks on infrastructure. In AI systems the two are fused, so attacking one often amounts to attacking both. Koyejo adds that risks track capabilities closely, so the more access and agency a system has, the more new security surfaces appear. He expects AI-native security companies to be more effective because they understand these systems deeply, and urges treating security and safety as model capabilities.