A new paper on arXiv warns that the apparent diversity of foundation models is deceptive. Beneath the surface, many models are built on a homogeneous technical stack of common training pipelines, distillation methods, and data. The authors describe this as a 'vulnerability monoculture': shared algorithmic choices that can expose otherwise distinct models to the same weaknesses.
The paper argues that this homogeneity has direct consequences for adversarial robustness. If an attacker finds an input that fools one model, the same attack may transfer to other models that share those underlying components. Evaluating models in isolation, therefore, misses a systemic risk that emerges across the ecosystem as a whole.
The authors propose a shift toward 'systemic adversarial robustness'—assessing how vulnerabilities propagate through shared infrastructure rather than treating each model as an independent target. The abstract does not provide experimental details, but the framing suggests that robustness benchmarks should account for the common substrate beneath many foundation models.