attacks

Definition

Inputs crafted to make a model fail or to bypass its constraints — adversarial perturbations in vision, prompt injection and jailbreaks in language. The recurring lesson is that defences are only ever evaluated against known attacks, so robustness claims expire. Most of these are not implementation bugs but consequences of what the model learned, which is why patching them one at a time rarely generalises.

Explainers using this term

• CPU Pipelines and Branch Prediction — The Factory Inside One Clock TickSpectre Attacks: Exploiting Speculative Execution
• Paper Walkthrough: Can Anything Catch a Fake Crisis Video? — What RA-Bench FoundCan We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? A Systematic Evaluation of Detectors, Generators and Social Dissemination