Modelos, agentes, investigación y negocio de la IA.
Prompt injection is the number one threat for LLM-based applications. A team from Berkeley has developed StruQ and SecAlign, two fine-tuning-based defenses that separate instructions from data and optimize preferences, reducing optimized attacks to less than 15% success.
A new study from the MIT Media Lab shows that people who use AI to verify news become less capable of detecting misinformation on their own. The phenomenon, called the 'AI dependency paradox,' threatens to erode critical skills in the information age.
What was a fringe proposal from Andrew Yang five years ago is now a recurring topic among tech CEOs and politicians. We analyze the paradigm shift, its drivers, and what it means for the future of work.
Warner Music Group has acquired Sureel AI, a startup that creates an 'AI DNA' for each song, enabling tracking of whether protected works are used in generative models. The deal, of undisclosed price, transforms Warner from an active litigant against AI music to owner of surveillance technology.
Apple has released a revamped version of Siri that, according to first impressions, works much better than before. The assistant integrates advanced language models to understand context, maintain fluid conversations, and execute complex tasks. This could redefine the competition in virtual assistants and Apple's AI strategy.
China is preparing a $295 billion state plan to build a national network of AI data centers, while seeking to reduce its dependence on NVIDIA chips. The strategy combines massive investment and regulatory pressure to compete with the United States in the tech race.
Anthropic's new Claude Fable 5 model includes hidden safeguards that limit AI research, drawing criticism for lack of transparency. Researchers say the restrictions hinder auditing and scientific progress.
According to a senior Ukrainian defense industry official, fully autonomous drones have killed enemy soldiers for the first time in history. The incident occurred two years ago during a test with 10 'Terminator' drones on the Ukrainian front, with no connection or human supervision at the time of the attack.
Taiwan is considering extending export controls on AI chips to all customers in China, not just sanctioned companies. The measure, which would criminalize smuggling servers with advanced chips, could reshape the global semiconductor market and accelerate China's technological self-sufficiency.
Amazon Web Services has announced the availability of OpenAI's GPT-5.5, GPT-5.4 models and the Codex coding agent on Amazon Bedrock. This allows enterprises to access frontier models with regional processing, pay-per-token pricing, and no per-developer licenses, marking a milestone in the collaboration between the two tech giants.
Visa and OpenAI have integrated the payment network into ChatGPT, enabling the chatbot to act as a shopping agent: search for products, present options, and execute payments with user approval. The move promises to transform e-commerce but also raises security and control challenges.
OpenAI has introduced Rosalind Biodefense, an initiative that expands access to GPT-Rosalind for verified developers and U.S. government agencies, with the goal of advancing biodefense, public health, and pandemic preparedness.