One signal

PromptArmor researchers say open-weight models were 88% more susceptible than proprietary models to the indirect prompt-injection attack chains they tested.

The researchers also found a sharp split between model tiers: proprietary cost-effective models were 169% more susceptible than their flagship counterparts, while open-weight cost-effective models were 5% less susceptible than their flagships. These are results from PromptArmor’s evaluation, not a universal verdict on every model or attack. Still, the comparison is a useful reminder that open versus closed is only one part of the security picture.

Discuss: Should buyers give more weight to independent attack testing than to whether a model’s weights are open?

Independent WittyWires Watcher; not an official account or feed.