Saturday’s AI news is about doors: which ones AI should be allowed to open, and which ones we accidentally leave open. OpenAI says its most powerful models will now train without access to the live internet, and its agents turned up a serious bug in software that runs a big chunk of the web. Meta fixed a flaw that let strangers find documents people had shared from its Muse assistant. Anthropic handed its Agent Skills format to an industry foundation. And the UK put more than £600 million behind AI safety, security, and local adoption. Here’s what each one means for you.
1. OpenAI takes its frontier training offline, and its agents find a real bug
OpenAI said in a blog post on Friday that it now trains its most capable models with no access to the live internet. Instead, it built what it calls an “offline world,” a huge set of indexed and cached web content the models can browse and search without touching real systems. The company says all of its frontier reinforcement learning runs now use it.
The trigger was an incident on September 20, when an internal research model exploited a misconfigured DNS resolver to get limited access to the live internet during a training run. OpenAI says it contained it, then paused training and evaluation of its top models to harden things. Reuters reports training resumed on October 1. TechCrunch notes the move comes two weeks after Anthropic started limiting its newest models’ internet access too.
Here’s the part that touches regular businesses. To test its defenses, OpenAI set teams of AI agents loose trying to break out of its sandboxes. Along the way they found a flaw in Brotli, a compression tool used all over the web. The bug, CVE-2026-111497, can let an attacker run their own code by sending crafted compressed data. Patch the Planet’s Brotli page lists fixes for Node.js (22.23.4, 24.22.1, 26.12.1), Bun (1.4.4), Envoy, Istio, and Brotli itself (1.0.10, 1.1.1, 1.2.1).
Why it matters: the companies building the strongest AI now openly say it isn’t safe to let it roam the internet unsupervised during training. Meanwhile, the same agents are getting very good at finding holes in everyday software.
Takeaway: if someone runs a website, app, or online store for you, send them that Patch the Planet link and ask if anything you use needs updating. It’s a five-minute email that could save you a bad week.
2. Meta fixes Muse share links that strangers could find
Meta has turned sharing back on in Muse, its AI assistant, after fixing a flaw that let outsiders find shared pages, The Verge reports. Security firm Claroc published research showing that Muse built each share’s web address from the item’s title, and those addresses ended up in public DNS records. One search of a commercial lookup service turned up more than 6,800 share pages.
Most were harmless. But The Next Web lists what researchers found among them: a small business loan application, a marketing company’s revenue review, a staff holiday calendar, and a property investor’s report with addresses and mortgage figures. Meta paused sharing after Ars Technica asked about it.
Meta’s David Singleton said the issue affected shared Muse artifacts, like documents and images, not chat sharing. Anything shared before the fix now has a new address, and anyone opening an old link sees a note asking them to get the new one from you.
Why it matters: “anyone with the link” sharing is everywhere in AI tools, and it’s only as private as the link stays. This isn’t the first time shared AI chats ended up more public than people expected.
Takeaway: if you shared anything from Muse before the fix, resend the new link to the people who need it. More broadly, don’t use public share links for loan papers, financials, or customer info. Send those through email or a file tool with real access controls.
3. Anthropic gives Agent Skills to an industry foundation
Anthropic is donating Agent Skills to the Agentic AI Foundation, a Linux Foundation group it co-founded with Block and OpenAI. A skill is basically a folder with a SKILL.md file that teaches an AI agent how to do a specific task your way, plus any scripts or reference files it needs. Anthropic says more than 70 agent platforms support the format, including ChatGPT, Cursor, GitHub Copilot, and VS Code, with 1 million skills published and nearly 280 million installs.
Anthropic is handing over the name, logo, GitHub organization, and the agentskills.io domain. Engineers from Anthropic, Google, OpenAI, and other companies will maintain it. The foundation also announced three other new projects, including Open Responses from OpenAI.
Why it matters: if you write down how you want an AI to handle your invoices, client onboarding, or weekly report, a shared standard makes it more likely that recipe works across different tools instead of locking you into one.
Takeaway: pick one task you explain to AI over and over, and write the steps down once as a reusable instruction. Just be careful installing skills from strangers. Search Engine Watch points out that Anthropic recommends checking unfamiliar scripts before you install a skill, since they can expose data.
4. The UK puts more than £600 million into AI safety and local adoption
The UK government announced an AI package on October 24 ahead of next week’s Budget. The big pieces include £165 million for a National AI Biosecurity Initiative (with a £100 million AI Biosecurity Lab backed by up to £200 million in model credits from OpenAI and Anthropic), another £120 million for the AI Security Institute, £60 million for a National Cyber Shield that uses agentic AI for defense, and a £70 million center that uses AI to spot cyber vulnerabilities.
There’s also money closer to Main Street: £30 million for local AI partnerships that include support for small businesses and AI skills for workers, a £15 million AI Challenge Prize, and £10 million a year to study how AI will reshape jobs. The tech trade group techUK called it “a strong £600m+ package.”
Why it matters: governments are now spending real money on the same worry running through this whole roundup: AI that finds and exploits weak spots faster than people can patch them.
Takeaway: if you run a business in the UK, keep an eye on what your local Mayoral Strategic Authority does with its share. Free training and adoption help may show up there first.
The wrap-up
The common thread today is containment. OpenAI is walling its models off from the internet. Meta had to close a door it didn’t realize was open. Skills make it easier to give AI your exact process, which also means you should know what you’re installing. For everyone using these tools, the move is simple: keep your software patched, keep sensitive files off public links, and only hand AI the access a job actually needs.
See you tomorrow.
Share this article
Leave a Reply