A security researcher ran adversarial tests against four AI orchestration frameworks (LangChain, CrewAI, AutoGen, SmolAgents) using the same underlying model, finding that framework choice alone produces a 2.6x difference in agent compromise rates (11.9% to 31.1%). The key insight is that orchestration frameworks make real architectural decisions — tool call validation, memory segmentation, autonomy controls — that directly shape the attack surface, independent of the model's own safety training. The post argues that security should be weighted equally with developer experience when selecting an orchestration framework, and that existing public comparison guides fail to include adversarial test data.
