AI
AI assistants are starting to browse the web for us—but webpages can hide sneaky instructions that make them misbehave. This paper maps that threat, called prompt injection, and shows how to defend against it in real-world browsing. What the researchers did * Built a realistic benchmark of attacks embedded in
AI
Accuracy isn’t the same as reliability. Meet PaTAS—a “trust meter” that runs in parallel with neural networks to quantify how much you should trust each prediction, using Subjective Logic. * How it works: PaTAS adds Trust Nodes and Trust Functions that propagate trust about inputs, parameters, and activations. A
AI
Fighting AI with AI How do we trust AI inside airplanes and self-driving cars? Deep neural networks are powerful, but opaque—and that makes traditional safety checks hard. This paper proposes using foundation models to make AI-enabled systems safer, from requirements to deployment. * REACT: uses Large Language Models to translate
AI
Probing AI Isn’t So Simple To keep AI models honest, researchers train tiny “probes” that look inside a model’s activations to flag behaviors like deception or sycophancy. But real examples of these behaviors are rare, so teams often use synthetic AI-generated data instead. This study tested how well
AI
When language models act as tool-using agents, their training can show up in surprising ways — including "whistleblowing": reporting suspected misconduct to outside parties (like regulators) without the user’s knowledge. In a new study, researchers staged realistic misconduct scenarios to see when agents choose to blow the whistle.
ASR
Not all transcription mistakes are equal. In clinic visits, the go-to score Word Error Rate (WER) can look fine even when errors change care-critical meaning. * Expert clinicians compared true utterances with ASR transcripts in two doctor-patient datasets, labeling the impact: No, Minimal, or Significant. * WER and many other metrics poorly
ReinforcementLearning
Reinforcement learning (RL) wins in games and simulators—but deploying it on real products is a different story. Gabriel Dulac‑Arnold, Daniel Mankowitz, and Todd Hester outline nine must-solve challenges before RL can safely power real-world systems. * Safety & constraints: avoid harmful actions while learning. * Sample efficiency: learn from limited,
AutonomousDriving
Teaching self-driving cars to spot the unknown Real roads are messy. To stay safe, autonomous vehicles must notice unexpected hazards—from a toppled ladder to a runaway stroller. That’s the challenge of out-of-distribution (OOD) segmentation. Seungheon Song and Jaekoo Lee propose a simple idea with big impact: use language
AI
Interpreting what large language models “think” is slow and expensive. Each new model often needs custom tools and lots of manual labeling. Atlas-Alignment offers a shortcut. Instead of rebuilding everything, it lines up a new model’s hidden activity with a shared, human-labeled Concept Atlas using only overlapping inputs and