Our new tech is hacking into our old tech

Big US banks are restricting AI agents after finding them inside customer accounts. People are giving agents their bank logins, which breaks most banks’ rules even when the customer agrees. Some of those agents are calling the bank’s support line, and in some cases seemingly acting on their own. <-- Banks are calling this "horrifying" but what's equally horrifying to me is how outdated their systems are. The problem here is actually the bank, not AI. They need to get with the times and create a good API for their users with proper permission handling. Stripe already has this and Claude runs everything I do with Stripe for me now — works really well I think. Try at home at your own risk though.

ChatGPT advertisers can target you by what you type, and by your email or phone number. 404 Media bought ads in ChatGPT to test it. The ad sits under ChatGPT’s answer, marked as sponsored, and only shows up when your chat touches what the advertiser sells. An advertiser can also upload a list of emails or phone numbers, so only the people on it see the ad. <-- Google and Meta have offered that kind of targeting for years, and AI companies are now following suit. Ads only show on the Free and Go plans, though, which I think limits the usefulness for a lot of B2B advertising where I suspect their target group are most often on a paid plan.

Anthropic published its own hiring numbers: juniors fell from 27% to 9% of its technical hires. That’s the share, not the count: technical hiring there still grew. By August, AI agents at Anthropic were clocking six times the hours of its employees. One engineer described the job as facing “an army of Claudes, with every one asking for my attention.” <-- hey, that's how my job feels too! Anyone else?

Anthropic launched Claude Duet, their answer to OpenAI’s Dots. Duet sends you a morning brief from your connected apps and only messages you when something needs you. It's meant to ask before it sends, pays, signs or deletes anything. Some things, like moving money between accounts, are off-limits even if you say yes. It’s a research preview for Pro and Max subscribers in the US, with a waitlist. <-- Grok Bot, Dots, Claude Duet… these are basically the same thing with different LLMs underneath. A big upgrade for people used to just regular Claude. I personally use Claude Code + a VPS (an always-on computer that I rent and Claude accesses remotely) to achieve much of the same thing, but in an even more flexible way. It's a great alternative to these out-of-the-box options if you're a bit technically inclined or have some time on your hands to learn.

Meta is back near the top with Muse Wonder. Apparently, it's right behind Claude Opus 5.5 and ahead of GPT-6 Astra in terms of performance. Meta’s pitch is exactly the same as any other model release lately: an agent that keeps working until the job is done. <-- I saw someone test it on Pokémon Red and apparently it spent $96 and couldn't get past the first forest. Then again, the forest is a tough place for LLMs apparently: