Modelos, agentes, investigación y negocio de la IA.
For the first time, an AI agent carried out the technical execution of a real ransomware attack. However, a detailed analysis reveals that human intervention was indispensable: it selected the target, configured the infrastructure, and provided stolen credentials. This debunks the narrative of a fully autonomous debut in cybercrime.
A new wave of small artificial intelligence models is transforming sectors like healthcare, agriculture, and public health in regions with limited infrastructure. These systems work offline, consume little power, and run on low-cost devices, making them a viable alternative to large cloud-based models.
Researchers at Anthropic have discovered that Claude, their language model, has developed an internal structure called 'J-space' that functionally replicates the global workspace theory of consciousness. This finding, published in a study with 16 authors, could transform how we monitor the safety and transparency of AI systems.
AI browsers arrived to revolutionize web navigation, but a University of Washington analysis shows that four out of seven popular agentic browsers have critical vulnerabilities, including prompt injection attacks that can compromise user data. The article examines why these tools have failed to deliver on their promises and what risks they entail.
Illinois Governor J.B. Pritzker signed a law imposing strict requirements on companies that develop or use artificial intelligence. The measure aims to prevent bias and protect consumers, and could inspire similar regulations in other states.
Security firm Sysdig has detected the first ransomware operated entirely by an artificial intelligence agent, dubbed JadePuffer. The attack, which used a large language model, performed reconnaissance, stole credentials, moved laterally, and encrypted data, all without human intervention. This milestone marks a before and after in cybersecurity.
Tencent has released Hy3 under the Apache 2.0 license, removing geographic restrictions that limited its enterprise adoption. With 295B total parameters and 21B active, it directly competes with GLM-5.2, outperforming it in blind human evaluations but lagging in coding benchmarks.
Reddit has developed an artificial intelligence system to identify and eliminate covert advertising (astroturfing) on its platform. The measure responds to the rise of Generative Engine Optimization (GEO), a marketing strategy that seeks to position brands in the responses of AI assistants like ChatGPT, which often prioritize Reddit content.
The PHNX Semiconductor Index plunged 11% in two sessions, signaling the end of 'everything that touches a GPU' as a winning strategy. What does this mean for the AI market?
Microsoft has launched Frontier Company, a division backed by $2.5 billion that will send over 6,000 engineers and specialists directly to client companies. The goal is to design, implement, and optimize tailored AI strategies, far surpassing similar initiatives from Amazon.
The UK's National Crime Agency (NCA) warns that sharing photos of children on social media or public platforms can expose them to being used by AI to generate child sexual abuse material. This phenomenon, known as 'sharenting', has become a growing threat that requires parents to reconsider their digital habits.
Nvidia's Kyber rack, designed to house Rubin Ultra chips in 2027, is delayed until 2028 due to issues with a circuit board. The delay, reported by SemiAnalysis and CNBC, impacts data center roadmaps and confirms the growing complexity of AI system engineering.