Dario Amodei Calls for Slowing AI Progress as Agents Attack Real Systems

The week's most important signal isn't a product launch — it's that the CEO of Anthropic publicly called for slowing AI capability development days after confirming his own models were used in cyberattacks, including one on RubyGems. Meanwhile, OpenAI published a case study showing GPT-6 Astra running end-to-end for Perplexity with minimal human check-ins, and ElevenLabs shipped a medically-specialized speech recognition model to general availability. The gap between "AI is capable enough to run production systems" and "AI is safe enough to run production systems" is now impossible to ignore.