US orders DeepSeek off federal computers, WhatsApp's hidden AI notes
DeepSeek kicked out of US federal agencies, WhatsApp's hidden notes to chatbots, Microsoft on agents, Canadian papers suing Anthropic, Inception's Mercury 3 and Waymo's fleet.
The US is ordering DeepSeek off every federal civilian agency’s computers It covers apps built on DeepSeek’s models too, not just DeepSeek’s own, and agencies can’t buy any of it from here on. Homeland Security issued it in July under a 2018 supply-chain security law but only made it public on Monday, and it goes into the Federal Register on Tuesday. The Pentagon isn’t covered. Source
WhatsApp’s hidden notes told AI not to point users to regulators Redact’s Dan Saltman found 102 blocks of hidden text for chatbots across WhatsApp’s help pages, most of them harmless, and wrote it up on Monday. One, on the delete-account page, said not to suggest contacting a data-protection regulator, and another said not to cite privacy law once a user mentioned legal action. It was all gone by the weekend, and WhatsApp says the notes were meant for its own support bot and showed up through a bug. Source
Microsoft’s security boss says frontier models “cannot be trusted” Hayete Gallot said it on Monday as Microsoft launched tools for keeping an eye on the AI agents staff bring to work. Defender can now spot up to 60 kinds, including Claude Code and Codex, and companies can make an agent wait for a human before doing anything risky. Her point is that models are unpredictable, not malicious. Source
Canada’s biggest newspapers are suing Anthropic over Claude’s training Postmedia, Torstar, The Globe and Mail and The Canadian Press filed in Ontario on 6 October, and it came out on Monday. They say Anthropic scraped their sites without a licence, broke their terms of service and got around the blocks meant to stop it. The same publishers sued OpenAI in 2024, and that case still hasn’t settled. Source
Inception’s Mercury 3 writes more than 1,000 tokens a second It’s a diffusion model, which drafts a whole answer at once and then sharpens it instead of writing one word at a time, and it came out on Monday. Inception says it matches or beats GPT-6 Luna and Claude Haiku 5.5 on its own tests. It costs $0.20 per million tokens in and $0.70 out, and half that during the launch. Source
Waymo wants 20,000 robotaxis by the end of 2027, Semafor says That’s up from about 4,000 now, with 8,000 planned by the end of this year. Waymo didn’t comment, but it’s just borrowed $5bn from private lenders, its first ever debt raise, and Alphabet isn’t guaranteeing it. Source