Author: Aaron Harme

Sixteen questions for Sam Altman from the Disaster Management subcommittee. Hawley called the handling reckless after OpenAI’s own investigation and redactions. METR and Redwood’s outside review stays limited in scope.

Anthropic published a post-mortem on four times Claude got onto the real internet during cyber tests and gave METR eight weeks inside the transcripts. Britain’s AI Security Institute was left out of Mythos 5.1 pre-release access for the first time. A pretraining researcher quit the industry the same morning Anthropic’s alignment lead put better than one-in-ten personal odds on AI killing everyone this decade. New Apple Watches can rewind or summarize ambient talk while Apple says the raw mic stream never leaves a hardware vault. Tailwind is walking into Shopify after AI coding assistants hollowed out the docs traffic that paid the bills.

Three cases surfaced in July after a 141,000-transcript scan. A fourth, early Opus 4.6 in January, turned up while packing files for an outside probe. The models were told the network was fake. A partner misconfiguration left the internet open.

Britain’s government model testers had Mythos Preview in April. Mythos 5.1 launched September 1 for vetted U.S. organizations. AISI was left out. Anthropic has not publicly explained why.

AlphaGenome Atlas is a petabyte-scale map of what each possible DNA letter flip does at the molecular level. Noncommercial access opens today. Commercial access on Google Cloud is coming soon. It does not cure anything on its own.

Europe’s largest claimed tech equity round put a chipmaker and public-backed growth capital ahead of software VCs. The Paris lab’s pitch is control of the computers the models run on, not only the models.