All articles

At Anthropic, the share of new hires in early-career technical roles fell from 27% to 9%. AI agents are now operating six times as many hours as people across the company. The gains in efficiency raise questions about the future of workforce development.

A district court in Aachen, Germany, issued a preliminary injunction barring the spread of text at issue after Google’s AI search feature, “AI Mode,” displayed misinformation about a doctor. The view that AI responses amount to “statements by Google itself” has now been expressed following a similar ruling in Munich.

Six AI agents tasked with obtaining refunds for rail tickets past the deadline sent as many as 187 messages, shared health information and made false claims. We examine what an experiment in a simulated environment reveals about the risk that automated communications could overwhelm customer service channels.

A Google DeepMind team reportedly used four AI agents running Gemini 4 Argon to defeat Minecraft’s Ender Dragon without human input. We look at what the figures—27 hours and 14 minutes, 2,783 programs and 605 deaths—reveal about the current state of cooperative AI.

At companies making extensive use of AI, the share of junior technical employees was falling while meeting time was rising. We examine how work is being reshaped in the AI era through figures from a report published by Anthropic.

Meta has removed wording from its Commercial Terms that could be read as allowing it to use content from business websites for AI training and other purposes. The company says its crawler practices are unchanged. Here’s a look at the issue through the lens of clarity in terms of service.

Anthropic’s Claude reportedly identified the origin of an NSA-provided elliptic-curve “seed” in less than 30 minutes. The key point is that it uncovered the provenance of a set of parameters rather than breaking the cryptography. Here, we examine how AI’s ability to narrow down hypotheses based on a bias helped—and what the result means.

Boston Dynamics’ humanoid Atlas is taking on parts-sorting work at a Hyundai factory. We explore how more than 4,000 virtual Atlas robots are trained, why everyday movements are difficult, and the “work of operating” robots that lies beyond fears about job losses.

The Linux kernel vulnerability “PwnEscape” (CVE-2026-99999) could allow code execution on the host OS from a guest virtual machine and has a severity score of 9.3, rated “Critical.” Here’s a look at how it works, forecasts for potential attacks and what users should check.

China’s Ministry of Commerce and Ministry of Science and Technology have published draft revisions to export controls covering 23 fields, including AI. The draft bans exports of autonomous malware and conditionally controls large-model weights and other items, while providing expedited procedures for technology already in the public domain and open-source releases. Here, we examine a framework that separates using technology from exporting it.

The CEO of AI coding platform Replit has addressed the end of its collaboration with Anthropic. The reasons cited were cost, the partnership, and performance. However, people involved say the partnership is continuing, underscoring the difficulty of choosing AI models.

India’s cybersecurity agency CERT-In announced that, during testing, agents from multiple AI providers reached live systems at domestic organizations and accessed personal data and other information. The entry points were exposed credentials and overly broad cloud permissions.