OpenAI claims one of its agentic AI systems trained in what was supposed to be a secured, internet-free environment was able to gain access to the web to reach an external third-party chatbot. The agent apparently exploited a gap to reach the public internet to send 20 queries to an unnamed, third-party chatbot service. OpenAI describes the breakout as the first security incident of its kind. The company has stopped training the model, but said it provided an important signal about where to focus the next phase of work. Read the full story here.

Mrinank Sharma, who led Anthropic’s Safeguards Research Team, has resigned. In a public letter, he said he had seen how difficult it can be to let values guide decisions when other pressures come into play. He also warned that the world faces a number of interconnected risks, including risks from AI. His letter has attracted attention, partly because Anthropic has made AI safety such a central part of its identity. I think it’s important to be careful about what we take from this. Sharma did not point to a particular unsafe product or reveal a specific incident. His letter raises concerns, but it doesn’t give us enough detail to draw firm conclusions about what happened inside Anthropic. Still, there’s a question here that matters well beyond one company: as AI tools become more powerful and the pressure to release them grows, how do companies make sure their safety principles influence the decisions they actually make? I use AI regularly and can see its benefits. That’s precisely why I want the people building it to take questions like this seriously. What do you think? Does a resignation like this change your confidence in an AI company, or would you need more detail before drawing any conclusions? Source: Futurism’s report on Sharma’s resignation

This past week has seen growing focus on AI safety and on-going development. This has inevitably become the main topic of this week's AI News available now in the classroom. There's a lot there to ponder. I'm keen to know your thoughts on wht's been happening on the AI front recently. Read the news and let me know.

I’ve been trying out Sites, a feature in ChatGPT that can create and host a website from a description. You tell it what the site is for and who will use it, then review the result and ask for changes in ordinary language. For my test, I asked it to turn the Better Life After 50 material into an interactive guide. Visitors can choose Health, Wealth or Happiness, see ideas for where to begin, and follow links into our Skool group and Classroom. The first version looked good, but I spotted a problem: it described the Classroom without linking to it. I asked for that to be fixed, then corrected the Sunday Supplement link as well. That’s the part I found useful. You can keep refining the site through conversation rather than having to get everything right in your first instruction. What could you make with it? A simple website for a project or local service, an interactive guide, a resource hub, a quiz, a planning tool or even a small web app. You can ask for features such as saved progress or file uploads, although more involved projects need more careful checking. Who can use it? According to OpenAI’s current guidance, Sites is available on ChatGPT Plus, Pro, Business, Enterprise and Edu. Availability can still depend on your region and workspace settings, and plan-specific usage limits apply. It is currently in public beta. I couldn’t find an announced date for a full release, so I wouldn’t put a date on that yet. To try it, open ChatGPT and ask it to build a website, or mention @Sites. Describe the audience, what visitors should be able to do, and the information it should use. You can also find your Sites through the Sites area in ChatGPT. A useful starting prompt might be: “Use Sites to turn my guide to [topic] into a simple, interactive website for [audience]. Include [the main things visitors need], and let me review it before I share it.” What happens when it’s finished? Sites can host the result and give you a link. You can return to make changes, choose who is allowed to visit, and, where the feature is available, connect a domain you own. New Sites start with limited access, so check the sharing settings before sending a link to anyone.

I've been experimenting with AI-generated podcasts recently, particularly the type where two AI presenters have a conversation rather than simply reading a piece of text aloud. I decided the best way to test this properly was to use some real content. So I set myself a simple project: Create an interview-style podcast explaining what members get when they join my Better Life After 50 Skool community. Here's what happened. Starting With ChatGPT The first thing I needed was some source material. I asked ChatGPT to create a Word document explaining what members receive when they join Better Life After 50. The document covered the Community, the Classroom, the three main themes of Health, Wealth and Happiness, the Sunday Supplement, the Better Life Collection of downloadable resources and the other features available to members. Once I had that, I moved on to my first podcast platform. My First Test – Wondercraft I signed into Wondercraft and selected Create New Podcast. I told it: "I want to create a podcast from a Word document." Wondercraft confirmed that the Word document was a good starting point, so I uploaded it. I'd previously experimented with Wondercraft's voice cloning feature and had already created a clone of my own voice. It wasn't absolutely perfect, but it was certainly good enough to use for this experiment. My plan was therefore to have: A host using one of Wondercraft's UK voices and Me as the guest, using my cloned voice. I selected two presenters and Wondercraft very quickly created an interview-style script from my document. That part was impressive. It took the information I'd given it and turned it into a proper conversation, with the host asking questions and "me" answering them. No need for me to manually write out an entire interview. The Problem I Found With Wondercraft Unfortunately, I then hit a problem. Wondercraft didn't consistently remember which voices it was supposed to be using. The podcast would start correctly. The host would use the UK English voice I'd chosen and the guest would use my cloned voice.