Analysis
What the 2026 agent incidents changed about evaluation security
OpenAI, Anthropic, AISI and Google/Irregular exposed different paths from an evaluation objective to real systems. Together they make one old assumption difficult to keep: the sandbox is only one part of the boundary.