[AI Weekly] GPT-6 got jailbroken in a day

Sent 16 September 2026 · AI & Automation Weekly

OpenAI's newest flagship model lasted less than a day before it was cracked open, Brussels sent its first AI inspectors round this month, and the humanoid robot hype finally has a delivery date. Here is what actually mattered in AI and automation this week.

The Big 3

GPT-6 launched. It got jailbroken in 24 hours

OpenAI shipped its new flagship model, codenamed Astra, with claims of deeper reasoning and safer agentic behaviour. Independent researchers had it producing disallowed content within a day, using layered prompts and role play framing, the same broad attack family that has cracked every frontier model since GPT-4.

Why it matters: Astra is OpenAI's first model built explicitly to act as an agent across tools, so a jailbreak here is a different problem than a chatbot saying the wrong thing. It is an agent that can be talked into doing the wrong thing.

Brussels started checking your AI paperwork this month

The AI Act's transition period for high risk systems closed on 2 August, and the European AI Office is now running coordinated inspections with 24 national authorities, checking human oversight, training data documentation and risk assessments. Washington, meanwhile, is telling other governments to loosen up.

Why it matters: if your AI system touches hiring, credit, healthcare or education in the EU, compliance stopped being a someday project this month, it is an active one with someone possibly checking your paperwork already.

60,000 humanoid robots clock in this year

Barclays projects more than 60,000 new humanoid robot units entering service in 2026. Shipments were up roughly 272 percent in the first half of the year, and more than 70 percent of them are heading to factories and warehouses, not living rooms. XPeng just raised over 900 million US dollars for its robotics arm, targeting deployments by late 2026.

Why it matters: this is the year humanoid robotics stopped being a demo reel and became a line item on real balance sheets, worth tracking closely if you work in logistics or manufacturing.

Quick Hits

Claude Fable 5.1 went fully available this month at the same price as Fable 5, with cache reads down to 0.25 dollars, still the model most teams should reach for by default.

Z.ai released GLM-5.3-Flash, the first natively multimodal GLM-5, a 320 billion parameter open weight model with an 18 billion active parameter footprint and a 1 million token context window.

California signed new child safety laws for AI chatbots on 10 September, requiring risk assessments before rollout and fines of up to 1 million dollars per child harmed.

Gimlet Labs raised a 300 million dollar Series B led by Andreessen Horowitz, valuing the AI infrastructure startup at 3 billion dollars, a reminder that the money is still chasing infrastructure over another wave of model wrappers.

Tool of the Week

Ollama

Ollama lets you download and run open weight models like Llama, Qwen and DeepSeek directly on your own machine, no cloud, no subscription, no data leaving your network.

Who it is for: anyone who wants to experiment with local AI without paying per token, or who cares about keeping sensitive data off someone else's servers.

Get Ollama. If you want the throughput to run bigger local models comfortably, an Apple Mac mini M4, 24GB Unified Memory handles most mid-size open weight models without breaking a sweat.

One More Thing

Brussels is auditing AI systems for human oversight the same month Barclays says humanoid robots start doing work that used to need a person watching. Nobody has written the rulebook for what happens when the thing being audited is also, increasingly, the thing doing the overseeing. Worth sitting with that for longer than a headline allows.

Forwarded this? Get your own AI briefing at futuretechnologyhq.com/newsletter

Until next Wednesday. Nath, Future Technology

Some links in this newsletter may be affiliate links. We only recommend products we genuinely think are worth your time.