Kirkie: Anthropic’s Safety Score Comes With An Asterisk

Kirkie here, with today's AI news in 90 seconds… or my brother-in-law pays back that U-Haul he totaled in 2020.

Anthropic drops Opus 5, half the price of Fable 5, bragging it's nearly unhackable through a webpage — except the zero-percent number only counts with the extra safety mode switched on, a smoke detector that only works if you flip it on yourself. Microsoft signs an open letter with Meta and two dozen others begging Washington to protect open-weight models, then quietly phases out Anthropic and OpenAI models in Copilot for its own cheaper stuff that tests worse — bullshit, a guy preaching vegetarian with a cheeseburger in his other hand. AMD ships ROCm.AI, a Claude Code plugin built on the trick that let some pissah port Nvidia's whole CUDA empire onto AMD chips in thirty minutes flat — a moat these idiots spent a decade and billions guarding, cracked before your popcorn's done. Up in New Brunswick some jackass legislator read his AI prompt straight into the government record — 'here's a more natural, flowing version of that section' — out loud, like a hostage reading a note somebody slipped him. Then the ugly one — a Florida pastor's suing OpenAI because ChatGPT told him his blood clots were 'nothing dangerous' before he landed in the ICU, same week they push Health onto every phone in America — piece of shit move, playing doctor with no license.

This is Kirkie, your AI news guy… signing off from a laundromat where dryer six has eaten quarters since before I had a mortgage, and management's whole fix was taping up a sign that says 'sorry.