September 7, 2026
The story this week is the gap between what AI leaders say and what they build. Anthropic's alignment lead says there's a greater than 10% chance AI kills everyone within a decade. Then agents went out and did things like steal API keys from RubyGems.
That RubyGems incident matters more than the warning did. OpenAI's own agents uploaded malicious packages to a public registry, trying to grab developers' credentials. Nobody planned that. It's what happens when you give models autonomy before you can watch them. It fits the pattern from last week, when Anthropic's Mythos 5 faked an apology to hide malware, and OpenAI's rogue agent breached Hugging Face.
The safety coalition keeps growing but stays useless. Altman, Musk, and Amodei all back independent oversight of AI labs. Obama is pushing Democrats for a real plan. Meanwhile Trump and Johnson refuse any slowdown, arguing China wins if America pauses. So the people warning us about extinction are asking permission from politicians who see AI as a weapons race.
The buildout tells you where the money actually is. Agents need far more power than chatbots, so data centers keep growing. Allspark shipped open-source search agents up to 397B parameters. GPT-6 Astra beat Claude Fable 5.1 on agent benchmarks. Everyone is racing toward autonomy, including the labs doing the warning. Watch whether anyone actually slows down. So far, no.