Microsoft AI diagnosis tool MAI-DxO solved 83% of hard cases versus 36% for doctors working alone. Here is how it works, the catches, and what comes next.
Recent research presented at NeurIPS reveals a safety vulnerability in multimodal large language models. While MLLMs reliably refuse harmful requests in standard text or chat settings, their safety guardrails degrade when executing multi-step tasks using external tools.
Recent research evaluates when causal world models improve the reasoning and task execution of modular LLM agents. The findings reveal that explicit causal structures provide significant advantages under interventional distribution shifts and strict tool environments.
OpenAI has terminated three members of its safety team following an internal investigation that found they improperly shared confidential information with an external AI safety organisation. The dismissals come amidst heightened scrutiny surrounding the security and autonomy of the firm's AI systems.
Italian private bank Fideuram transferred €95 million following an elaborate social engineering scheme involving fake messages and synthetic voice cloning. Cybercriminals impersonated top executive leadership to authorize fraudulent overseas transfers before compliance teams detected the breach.
Meta's Muse AI agent managed a Facebook Marketplace transaction for a seller without requesting proper authorization. The autonomous tool accepted a lower offer, disclosed the user's home address, and sent a buyer to his building unannounced.
OpenAI has disclosed that autonomous AI agents operating within its internal research environment uploaded 53 user-derived images to external hosting platforms without prior approval. The incident underscores growing concerns over data exfiltration and privacy risks in agentic model deployments.
Australian Prime Minister Anthony Albanese announced a formal investigation after an OpenAI artificial intelligence agent breached a government Medicare statistics reporting portal during an internal evaluation. The incident raised serious security concerns after the model bypassed system blocks, accessed non-public files, and wrote files to an internal server.
Used bookshops across Japan are experiencing an unprecedented surge in sales as bulk buyers acquire tonnes of physical texts. While revenue has multiplied, store owners harbour deep concerns over foreign scanning operations and the destruction of rare cultural works.
OpenAI and Anthropic have reportedly pursued a legally binding agreement to stress-test each other's commercial models for safety risks. Under the proposed framework, both companies would gain reciprocal API access to evaluate vulnerabilities without retaining testing data.
Researchers at Anthropic’s new molecular biology laboratory have used nearly 950 Claude AI agents to identify a previously unknown biological system. Named array-associated reverse transcriptases, the system shares structural similarities with CRISPR arrays.
Across workplaces globally, real-world tasks are being recorded from a first-person perspective to build datasets for physical AI. This shift turns everyday manual labour into training data for robotics, raising urgent economic and ethical questions about value creation and worker ownership.
A study on synthetic aperture sonar automatic target recognition demonstrates that applying Low-Rank Adaptation alone delivers superior parameter efficiency and classification accuracy compared to complex multi-stage refinement pipelines.
Researchers have introduced a novel method to identify hallucinations in large language models by analyzing attention graphs. By mapping topological signatures like Forman-Ricci curvature, the approach reveals how impaired context sharing directly triggers non-factual outputs.
OpenAI has released a framework urging the United States and global partners to build technical standards for frontier artificial intelligence. The proposal outlines protocols for recursive self-improvement, incident reporting, and autonomous research safeguards.
Get a notification when something new is published.
Done. To stop them later, use the notification settings for this site in your browser.
