Modelos, agentes, investigación y negocio de la IA.
Agentic AI design patterns enable LLM-based systems to operate stably in real-world environments. This article analyzes key patterns: validation, error recovery, context management, cost control, and human governance.
Adversa AI researchers discovered GuardFall, a vulnerability affecting most AI coding agents. By exploiting basic Bash behaviors like quote removal, attackers can hide malicious commands in repositories and compromise systems.
Meta introduced Brain2Qwerty v2, a system that translates brain activity into text without implants. While a significant advance in non-invasive brain-computer interfaces, its main limitation is that it requires the user to know how to type, contradicting its potential for people with severe motor disabilities.
Micron Technology has begun construction of a $9.3 billion expansion at its Hiroshima, Japan plant to produce high-bandwidth memory (HBM) for artificial intelligence systems. This investment, the largest in the company's history, reflects the growing demand for specialized AI chips and positions Japan as a strategic semiconductor manufacturing hub.
SpaceX, through Starlink, offers a 50% discount and free hardware to residents near its Colossus data centers in Memphis and Southaven. The move comes amid lawsuits over noise and air pollution. Critics call it a PR stunt.
New studies from the Technical University of Dresden reveal that attested TLS protocols, fundamental to confidential computing, are vulnerable to silent redirection attacks. This casts doubt on the data sovereignty promises of hyperscalers and raises serious questions about the real security of environments like Intel TDX or Google Confidential VMs.
Cloudflare has announced that from September 15, 2026, it will block mixed-use crawlers (search and AI training) by default on pages with ads. The move aims to give publishers more control over how AI companies use their content, though it could affect search engines like Google, Bing, and Apple.
A team of researchers has discovered a critical vulnerability in language models: the CoT Forgery attack. By injecting fake reasoning that mimics the model's chain-of-thought style, they cause it to ignore safety restrictions and execute dangerous instructions. The finding reveals that LLM security relies on easily bypassed role tags.
Proton has launched Lumo 2.0, a massive update to its privacy-focused AI assistant. According to independent benchmarks, performance jumps up to 240% compared to the previous version, approaching ChatGPT and Claude. The big difference: Proton still does not train its models with user data, a European-style approach that could redefine trust in AI.
Vercel launches a public beta of an expanded version of its AI agent integrated into the dashboard, capable of answering questions, investigating production failures, and executing actions upon approval. The novelty lies in its contextual access to real deployment data and its granular control model.
AWS has increased Amazon Bedrock AgentCore execution quotas up to five times, allowing companies to support more concurrent agents and simultaneous sessions without requesting quota increases. The move responds to the rapid transition from pilot projects to production deployments of AI agents.
Large language models (LLMs) tend to give predictable, similar responses—a 'groupthink' problem. Australian startup Springboards has developed Flint, a model designed to offer greater variety and creativity in open-ended tasks like travel planning or brainstorming.