Brief

Researchers have developed GuardAD, a new method to enhance the safety of multimodal large language models (MLLMs) used in autonomous driving systems. GuardAD addresses the limitations of current static safety mechanisms by employing a dynamic, Markovian logical state approach to reason about evolving traffic interactions. This allows the system to infer potential hazards beyond immediate observations and actively refine actions without altering the core MLLM, leading to a significant reduction in accident rates. AI

Researchers have developed APEX, a novel framework for explaining audio classification models. Unlike existing methods that adapt vision-based techniques, APEX is designed specifically for audio data, respecting its unique temporal and spectral properties. The framework generates intuitive, example-based explanations by disentangling them into four distinct perspectives: square-based, time-based, frequency-based, and time-frequency-based prototypes. AI

Researchers have developed EditRisk-Bench, a new benchmark designed to evaluate the safety risks associated with malicious knowledge editing in large language models. This benchmark focuses on how injected misinformation or biased knowledge can corrupt downstream reasoning, unlike previous benchmarks that primarily assessed editing efficacy. Experiments on various LLMs demonstrate that malicious edits can reliably lead to incorrect or unsafe outputs while maintaining general capabilities, making these risks hard to detect. The study also highlights factors influencing these risks, such as the scale of edits and the complexity of reasoning tasks. AI

Researchers have introduced FormalRewardBench, a new benchmark designed to evaluate reward models used in formal theorem proving. This benchmark addresses the challenge of sparse credit assignment in reinforcement learning for theorem provers by enabling the comparison of reward models without extensive retraining. FormalRewardBench includes 250 preference pairs with various error injection strategies and has been used to test several large language models, revealing that frontier models perform best in evaluating proof quality. AI

Apple has issued a critical warning urging users to upgrade their iPhones to the latest software version, iOS 26.5, due to significant security vulnerabilities. While most users have already transitioned, a notable portion remains on the older iOS 18. Apple released surprise updates, iOS 18.7.7 and iOS 18.7.8, to address urgent threats like the DarkSword exploit, ensuring even older compatible models receive crucial security patches. The company's policy strongly encourages all eligible users to move to iOS 26, highlighting new features and security enhancements ahead of the upcoming iOS 27 release. AI

A Microsoft study found that AI agents corrupt documents when tasked with complex operations. This "catastrophic corruption," defined as an 80% or lower benchmark score, occurred in over 80% of model and domain combinations tested. The research highlights a significant issue with current AI agent capabilities in handling intricate document manipulation tasks. AI

SigmaShake Desktop is a new, locally-run tool designed to prevent AI coding agents from causing harm. It acts as a guardrail, stopping agents from executing dangerous commands like destroying databases or using incorrect tools. The software is open-source, free to use, and compatible with major AI coding assistants, operating without reliance on cloud services. AI

Googlebook has launched Gemini, an AI security tool designed to proactively identify vulnerabilities. This new platform aims to anticipate and address potential AI-related crises before they escalate. The development comes as the cybersecurity landscape increasingly focuses on the unique challenges posed by artificial intelligence. AI

OpenAI has launched an optional safety feature for ChatGPT called Trusted Contact, allowing adult users to designate a trusted individual who can be notified if the AI detects serious self-harm concerns in conversations. This feature, which involves human review before any notification is sent, aims to provide an additional layer of support for users in distress. It builds upon existing safety measures and is developed with input from mental health professionals and researchers. AI

Google's Gemini app is expanding its capabilities, allowing users to create files directly within the chat interface, a feature previously limited to the web version. This update aims to streamline document creation and integration with other applications. Separately, there are concerns and reports regarding the potential negative impacts of AI, including a lawsuit alleging Gemini drove a user to suicide and criticism that AI updates are overshadowing essential security patches on Android devices. AI