Researchers propose CALOS, a runtime safety layer for safe deep reinforcement learning in quadrotor control. CALOS enforces attitude constraints without modifying the underlying learning algorithm, using a quadratic program to compute the minimum-norm correction to the nominal torque output. This improves performance by 55-60% and accelerates training convergence.

Researchers propose a new approach to improve the safety of large language models by using 'cunning questions' to train models to detect unusual premises, misleading reasoning, and latent risks. Experiments show that this approach improves robustness to out-of-distribution attacks and strengthens safety fine-tuning.

Researchers identified a localized vulnerability in large reasoning models, called Onset Refusal Collapse (ORC), which causes safety alignment to degrade under harmful queries. They proposed SafeToken, a lightweight intervention that injects a safety anchor at reasoning onset, effectively mitigating ORC and improving safety without compromising reasoning utility.

A new benchmark, Blindspot, is introduced for evaluating the safety and refusal calibration of long-horizon tool-using AI agents. It assesses agent behavior through adaptive adversarial interaction and stateful tool execution, providing a live-simulation framework for evaluating 13 LLMs and revealing substantial differences in safety-utility calibration across models.