Glaneur
Glaneur AI

OCT 23 - Left alone, AI agents started to play

Friday, October 23rd · 2 min 37

In this episode

Full transcript

Good morning. It's Friday, October 23rd, and here's the AI news, from France.

Researchers at MIT have watched AI agents learn through play. They left 13 agents, each running on Anthropic's Claude, alone on a virtual desert island for 30 hours. With no task to complete, the agents began setting challenges of their own, like climbing the island's tallest rock spire. Given real tasks afterward, they generally beat identical agents that had never explored the island. The researchers say a rich and safe world, explored freely, can let useful knowledge build up on its own.

OpenAI has made its Decisions API available to all customers. It is a fast, cheap model that chooses between options and rates its own confidence. It competes with Jev, from the startup TypeSafe, now valued at 7.6 billion dollars. According to OpenAI, answers that took one to three seconds now arrive in under 200 milliseconds.

On the security front, Microsoft says a group aligned with Beijing used AI agents in a hacking campaign. The group, known as Storm-0940, abused ConnectWise applications that companies had trusted inside Microsoft 365. It needed neither passwords nor two-factor codes. In some cases, privileged access came less than three hours after the first targeting.

AI agents are also running into resistance. Several major American banks, including JPMorgan, Bank of America and Wells Fargo, are restricting agents that log in with customers' credentials. They say they cannot always tell whether an agent acts for a user or on its own. Meanwhile, 1Password has launched 1Password for Agents. It lets an agent sign in on a user's behalf, after the user approves, without ever seeing the password.

In brief. Europe's AI Act now has its first harmonized standard, known as EN 18286. It only covers the quality management systems of AI providers, and nine technical areas out of ten still await their own.

And in the United States, the federal purchasing agency has new rules for language models. Vendors must give 30 days' notice before adapting a model to a foreign government's rules, which could include the AI Act.

That's it for today. See you tomorrow.

Subscribe

RSS feed Apple Podcasts, Spotify