The Security Model That's Getting More Capable Faster Than Its Makers Expected
The UK's AI Security Institute tested a new checkpoint of Anthropic's restricted Mythos model and found it had already surpassed its previous results, solving a cyber test range no AI had cracked before. AI cyber task capability has been doubling every 4.7 months, and the benchmarks may already be too small to measure what these models can actually do.
Read full story →For the First Time, More Businesses Pay for Anthropic Than OpenAI
Real spending data from 50,000 US companies shows Anthropic at 34.4% enterprise adoption versus OpenAI at 32.3%, the first time Anthropic has led. Claude Code drove the shift, a PwC deal will train 30,000 staff, and OpenAI is responding by reinventing itself as an infrastructure and consulting firm.
Read full story →Every AI Prompt Has an Energy Cost. Almost Nobody Is Counting.
Microsoft has quietly retreated from its 2030 carbon-negative pledge as AI infrastructure drives energy demand sharply higher. Most organisations deploying AI have no visibility into the electricity consumed per query. The efficiency case for AI is harder to make when the energy bill is invisible.
Read full story →The Worm That Hid in Your Build Cache and Impersonated Claude
A supply chain attack dubbed Mini Shai-Hulud compromised 169 npm packages and 2 PyPI packages, including the official Mistral AI client and TanStack's react-router (212 million monthly downloads). The malware stole cloud credentials, spread as a worm, and disguised itself by impersonating Anthropic's Claude bot in repository logs.
Read full story →