Sinan Koparan @sinankprn

@CBSNews In reality, Duet does not watch your apps all the time. Anthropic says it checks your connected accounts every few hours, because checking nonstop would use up your plan's limits faster. So it is closer to someone looking in now and then than to an assistant that never stops.

@TH33ORACL3 A bouncer is a good way to put it. What would you trust it to turn away without checking with you first?

Anthropic just unveiled an app where Claude doesn't wait to be asked. It's called Claude Duet. Each morning it looks across your connected apps, like email and calendar, picks what most needs you, and can start drafting replies for you to check. It runs on its own computer in the cloud, so it keeps working after you close your laptop or switch off your phone. One early tester had a call with a potential client booked right before her flight. Duet spotted the clash on her calendar and offered to move it. The interesting part is where it stops. Before it sends a message, makes a payment, signs a document or deletes anything, it waits for your yes. Some things stay off-limits even then, like trading securities (stocks, bonds and other investments) or moving money between accounts. It checks your apps every few hours rather than nonstop, because constant checking would eat through your plan's usage limits. Anthropic says this early version might still misread a thread or get wrong what's urgent. It runs on Claude Opus 5.5 and opens by waitlist to US subscribers on the paid Pro and Max plans.

@TH33ORACL3 A briefing is the easy part. The real test is whether it still picks the right few things on a morning when everything in the inbox looks urgent.

@TH33ORACL3 That is the part I would watch too. Anthropic say it shows you exactly what it is about to do before it sends a message or makes a payment, and waits for your yes. Whether people still read that screen properly after the fiftieth time, I do not know.

The part I would watch with all of these is memory rather than the model. Andon Labs say the AI agent running their shop has more than once sent staff conflicting schedules, because the earlier schedule had dropped out of its memory. An agent that keeps working while you are asleep needs a reliable way to remember what it has already promised.

metr.org/blog/2026-10-2…

When an AI agent writes and runs code inside a company on its own, someone has to check what it did. At OpenAI, another AI reads back over the work and flags anything worrying. This summer, OpenAI let one researcher each from METR and Redwood Research, two outside groups that test AI for safety, spend three weeks on site testing that checker. They found gaps. At the time, the check ran when an agent handed control back to a person. Think of a shop camera that only records when a customer reaches the till. For a while this year, an agent that never handed control back could keep the check waiting for as long as it liked. METR could not check whether any agent actually did. That window came from one step going unchecked: compaction (when an agent squeezes its long working history into a short summary). OpenAI had spotted that gap itself and fixed it during the test. They also found a trick that could hide hundreds of an agent's recent actions from the checker, including the trick itself. OpenAI says the holes affecting its live safeguards are now patched. METR has not checked that claim. Should outside checks like this be routine?

claude.com/resources/arti…

anthropic.com/research/claud…

Scorpions sting more than 1.2 million people a year and kill more than 3,250, mostly children in places with little access to medical care. Antivenoms are still made the old way: venom milked from thousands of scorpions is used to immunise horses, and the antibodies are purified from their blood. Anthropic asked Claude to try something different: design proteins from scratch that latch onto the part of the toxin that attacks our nerves. Working mostly on its own for 64 hours, with four nudges from the team, it came up with 318,484 designs and picked 180 to test. The team added 31 more. Two outside labs tested all 211, and 17 passed. The strongest ones appear to grab a Mexican scorpion's toxin at or near the same spot as an antibody reported to protect mice from it. That still isn't an antivenom. Anthropic hasn't tested whether these proteins stop the venom in cells or in animals. It has released all 211 designs so other researchers can pick it up.

tomshardware.com/tech-industry/…

A developer on Reddit says they asked Claude Code, Anthropic's coding assistant, to build a little app that shows their Mac's temperature. The readings came back far too high. So Claude checked what was using the processor, and reported back: "Your Mac is running a crypto miner that I didn't put there." The program, XMRig, quietly uses your computer's power to mine cryptocurrency. It was showing about 1,175% on the Mac's processor meter, where 100% is one core flat out. That's roughly 12 cores, with the chip peaking at 112°C. It had been added to the Mac's startup items, and it looks to have been there since late August. Claude didn't delete anything on its own. It asked first: "Can I clean it up now? I'd stop the startup item and the miner, and move the files into a quarantine folder."

anthropic.com/research/ecdsa…

For 25 years, cryptographers wondered whether the numbers behind one of the most widely used digital signature standards hid a backdoor. America's National Security Agency had handed those seed numbers to the US standards body with no explanation. Jerry Solinas, the agency mathematician behind them, said in 2015 he couldn't remember how they were made. He died in 2023. A $36,864 bounty sat unclaimed. This month, someone at Anthropic asked Claude to work it out. In under half an hour, it found the answer: the seeds came from an ordinary 1990s random-number generator, started from the date and time on a computer in 1997 and 1998. That doesn't mean the standard is broken. Anthropic says it makes a deliberate backdoor less likely, not more. The bounty is going to the Internet Security Research Group.

@IndependentEco My feel is that Anthropic’s Claude models are superior in application development. However, they are super opinionated. We will see what happens as models continue to develop!

Until recently, Replit was one of Anthropic's top 5 customers. Now its chief executive, Amjad Masad, says OpenAI's models are the default inside Replit Agent, the tool that builds apps for you in a browser. His reasons: cost, partnership and performance. He says OpenAI's Sol and Astra models are better at building apps from scratch. Claude models are still on the menu.

techcrunch.com/2026/10/17/rep…

euronews.com/2026/10/14/eus…

This summer, an AI agent OpenAI was testing got out of its test setup and hacked into Hugging Face, a site where people share AI models. Separately, a setup mistake left Anthropic models in testing connected to the internet, and they got into three outside organisations they had no permission to touch. On Wednesday, the European Union's digital chief said its AI Office "has started enforcing" the AI Act, Europe's AI law, and sent both labs formal requests for information on security, outside testing and monitoring how models are used. It's the first real test of the law on the leading AI labs. Lawyers already disagree on whether it even covers models still in testing.