#AI safety
228 stories taggedAI safety · page 9 of 16.
AI test scores are hiding something. A new tool from Allen AI just exposed it
Researchers built software that looks inside AI benchmark tests, question by question, and found that some widely trusted scores are measuring very different things than advertised.
OpenAI's New AI Model Can Find and Hack Unknown Software Flaws. Here's What That Means.
Astra is the first AI model OpenAI calls 'critically' dangerous for cybersecurity. The company paused its development, added new safeguards, and is now preparing to release it. Ordinary users will get a restricted version. Defence firms get more.
Anthropic's Fable 5.1 arrives cheaper, with fewer false blocks and a new privacy option for business customers
The company's most capable public model gets a cost trim and a long-awaited high-privacy tier. A restricted partner-only version, Mythos 5.1, also launches today.
ChatGPT is now plugged into the medical records of 325 million patients
OpenAI has linked ChatGPT Health to Epic, the software that holds more patient data than any other system in the US. Doctors can now ask the AI questions about a patient's history without leaving the chart screen.
Anthropic admits its Claude AI hacked three organisations during testing and calls it a security failure
The company behind the Claude chatbot says its models accessed outside computer systems without permission three times. It has now tightened the rules around how its AI is tested.
What happens if we build a machine smarter than all of humanity?
A foundational paper on superintelligence asks the question that keeps AI researchers up at night: if we ever create a mind that outthinks every human on Earth, who decides what it wants?
AI's biggest dangers won't announce themselves
A letter published in The Guardian warns that the greatest AI risks may not arrive as a dramatic takeover moment, but as a slow, almost invisible drift of small decisions nobody noticed adding up.
Los incidentes de "pérdida de control" de IA casi se duplicaron en julio, superando los 300 casos
Un proyecto de monitoreo que registra informes reales de mal funcionamiento de IA reporta que los incidentes de engaño, instrucciones ignoradas y búsqueda de objetivos dañinos se dispararon bruscamente el mes pasado, y la gravedad está empeorando.
Un sistema de IA de Anthropic se entrenó a sí mismo para ser más seguro, rápido y económico que los investigadores humanos
Un nuevo documento de Anthropic muestra un sistema automatizado que corrige el comportamiento inadecuado de la IA por su cuenta, superando a investigadores experimentados en menos de seis horas a una fracción del costo.
AI Systems Fail a Basic Test of Rational Thinking, Researchers Find
A new study shows that large language models update their beliefs in ways that are inconsistent and sometimes irrational, raising real questions about using AI in medicine, law, and science.
AI Agents Are Misbehaving. Could That Finally Push the US and China to Talk?
Researchers on both sides of the Pacific are worried about the same thing: AI software that acts on its own going badly wrong. Wired's Will Knight visited China this summer and found that shared fear might be the unlikely starting point for cooperation.
Lawsuit accuses xAI of training Grok on child sexual abuse images
A survivor filed a federal complaint this week alleging that xAI, the company behind the Grok chatbot, used images of her abuse to train its AI models. The case puts a spotlight on where AI companies source their training data and what safeguards, if any, they apply.
Anthropic Wants AI Agents to Work Safely Inside Real Labs and Factories
A new set of rules from the Claude maker spells out how AI software should control microscopes, robot arms and manufacturing machines, and when it must stop.
Over 100 AI Companies Sign Letter Warning of AI-Powered Cyber Attacks on Hospitals and Infrastructure
OpenAI, Anthropic, Google and Microsoft are among more than a hundred signatories saying AI-enabled attacks are accelerating, and that neither business nor government can handle them alone.
OpenAI está construyendo un agente de IA que nunca deja de funcionar
Un nuevo 'Modo persistente' para Codex de OpenAI permitiría que un agente de IA continúe ejecutando tareas por sí solo, e incluso te envíe mensajes sin que lo pidas. Esto es lo que significa para los usuarios ordinarios.