Researchers have developed a new type of attack on visual world-model agents, such as DreamerV3. These attacks, called Critic-Induced Value-Subspace Attacks (CIVA), exploit the way these agents use a recurrent latent state to make decisions. CIVA works by first probing the agent offline with critic-guided perturbations and then extracting a low-dimensional subspace of the agent's value function. At test time, it optimizes only the coefficients of this subspace, keeping online optimization cheap and temporally coherent. The authors claim that CIVA consistently outperforms five recent methods in various experiments. --- Why it matters: This matters to researchers in AI because it shows a new vulnerability in visual world-model agents, which are increasingly being used in applications such as robotics and game playing. Understanding this vulnerability can help improve the security of these systems. Source: https://arxiv.org/abs/2608.21114