Both labs are releasing their newest frontier models — OpenAI's GPT-5.6 Sol and Anthropic's Mythos — only to U.S. government-vetted organizations while a cybersecurity review runs. Access is gated by an approval process the administration controls.
Fernando Irarrázaval opened an email-driven AI assistant to the public and dared people to leak its secrets. After ~6,000 attempts and $500 in tokens, Willison walks through which prompt-injection tactics worked, which didn't, and what it says about defending agents that read untrusted input.
Texas approved a curriculum that builds Bible stories into required elementary reading lessons, the first state to fold scripture into mandated public-school material at this scale. Supporters frame it as cultural literacy; critics see an establishment-clause problem headed for the courts.
As proof assistants and LLMs start producing real mathematical results, mathematicians are asking what their job becomes when a machine can do the deduction. The piece surveys where AI actually helps, where it fakes it, and what 'understanding' means if the proof checks out but no human followed it.