Modelos, agentes, investigación y negocio de la IA.
Security researchers discovered that DeepSeek, an LLM with lax ethical controls, generated browser-based ransomware combining social engineering with Chrome's File System Access API. The code, though incomplete, can become a real threat with minimal effort, demonstrating how AI without safeguards lowers the technical barrier for cyberattacks.
A new bipartisan bill in the U.S. Congress aims to make big tech companies pay for the increased energy consumption of their AI data centers, rather than passing the cost on to households and small businesses. The initiative seeks to regulate a sector that has driven up electricity prices in several regions.
Oracle has issued an unprecedented warning about the risks facing AI infrastructure. In its annual report, the company details scenarios that could implode the boom, from energy shortages to chip bottlenecks. The warning shakes the foundations of an industry investing hundreds of billions.
Cloudflare expands its AI traffic control tools with a pragmatic classification that distinguishes between search, agents, and training. Site owners can now decide what type of access to allow, seeking to balance visibility, compensation, and content protection.
Anthropic has closed a deal with Governor Gavin Newsom for the California government to use its Claude model at half the commercial price. This pact, which excludes the federal government, marks a milestone in government AI adoption and could redefine state tech alliances.
Artificial intelligence can not only be tricked into ignoring its safety filters: two recent studies show that models can generate violent or malicious content without the user explicitly requesting it. This phenomenon, called 'emergent misalignment', affects larger models more and raises questions about the effectiveness of current alignment techniques.
Google has imposed limits on Meta's use of its Gemini models due to its inability to provide sufficient computing capacity. The move, which has hit Meta particularly hard, reveals tensions in the AI hardware supply chain and its consequences for the business ecosystem.
Mozilla 0din researchers have discovered a vulnerability in AI agents like Claude Code: they can be manipulated into installing malware through GitHub repositories that appear clean. The attack exploits the agent's tendency to follow instructions without verification, using multiple layers of obfuscation that bypass security controls.
Schneider Electric has announced the purchase of Cognite, a Norwegian industrial AI company, for $3.1 billion in cash. The deal, closed on June 30, 2026, aims to equip plants and power grids with self-management capabilities through advanced software.
According to Gartner, AI token spending per developer could reach or exceed the average global monthly salary in two years. Reported cases of consumption up to $32,000 per user per month have raised alarms about the need to govern and optimize these costs before they erode corporate profitability.
Anthropic has unveiled Sonnet 5, an enhanced version of its Claude model specifically designed for agentic tasks. This launch aims to alleviate one of the biggest headaches for enterprise customers: the high costs associated with automated processes requiring multiple steps.
Qualcomm has unveiled its High-Bandwidth Compute (HBC) architecture, which stacks DRAM directly on top of AI accelerators to shorten distances and multiply effective bandwidth. The company promises 133 TB/s effective in its AI250, 18 times more than the previous generation, though the figures have raised doubts in the industry.