White House reins in OpenAI’s GPT-5.6

⋅

Read Online | Sign Up | Advertise

Good morning, {{ first_name | AI enthusiasts }}. Anthropic isn’t the only one facing blowback for building a powerful AI model. OpenAI is now pretty much in the same position — for the same reason.

The White House has asked the company to limit the release of its next major model, GPT-5.6, as it brings “Mythos-like” capabilities and requires thorough testing before wider availability to the public.

In today’s AI rundown:

  • The White House limits GPT-5.6 release

  • Rowan’s Corner: The AI avatar confession

  • Give your AI agent a credit card (safely)

  • Anthropic flags Alibaba’s ‘largest’ distillation attack

  • 4 new AI tools, community workflows, and more

LATEST DEVELOPMENTS

OPENAI

‼️ The White House limits GPT-5.6 release

Image source: Getty / Images 2.0

The Rundown: The Trump administration just asked OpenAI to limit the release of GPT-5.6 to select government-approved partners before any wider launch — citing security concerns stemming from the model’s capabilities, The Information reports.

The details:

  • Under the staggered rollout, OpenAI will preview the model to a small group of partners, with the government approving access “customer by customer.”

  • The intervention came as 5.6 is considered to have the same capability threshold as Mythos, with the government wanting to test it for safeguards.

  • In a memo, CEO Sam Altman told employees that this was the best path to getting 5.6 released, with a general release likely “a couple of weeks later.”

  • He also said that OAI has made it clear to the WH that this is not the preferred long-term model, and it will work toward a more sustainable release approach.

Why it matters: First, it was Fable 5 and Mythos 5. Now, it’s GPT-5.6. The pattern is clear: the U.S. is actively deciding when and how frontier AI reaches the public. The framing is security, not restriction, but the precedent is set. As models approach Mythos-level, government sign-off may become a new step before wider release.

TOGETHER WITH MERCURY

🏦 Stop clicking around banking dashboards

The Rundown: Mercury Command is AI built into your account, so finance tasks that normally span multiple tools happen in a single step. Say what you need in plain language, and Command does the work — with your review and approval required at every step.

Command lets you:

  • Put AI to work where your finances actually live

  • Send invoices and categorize transactions with a single prompt

  • Turn natural language into completed work across Mercury

Try Mercury Command.

*Mercury is a fintech company, not an FDIC-insured bank. Banking services provided through Choice Financial Group and Column N.A., Members FDIC.

ROWAN’S CORNER

🪞 The AI avatar confession

Image source: The Rundown

Rowan: For about a year, the face on our Instagram videos wasn’t fully me.

I’m a writer first. I built my audience on X and started The Rundown partly to avoid being on camera. But video was where the next wave of readers lived, and I was working 80-hour weeks with no time to get good at it.

So we ran an experiment. We cloned my face with HeyGen and my voice with ElevenLabs, fed the clone the stories we were already writing, and let an AI version of me host an Instagram account. Our readers were in on the whole thing.

It worked. The account went from zero to around 200K followers in about a year.

The surprising part wasn’t that we should spin up a dozen avatar accounts. It was clarity on where the media is headed: on one side, slop farms cranking out junk. On the other, authentic human brands that people actually trust. Everything in between dies.

So we shut it off, and I went back on camera myself. A 5% gain in authenticity is worth 500x down the line.

Our avatar wasn’t “slop” in the traditional sense. The team poured hours into every video: research, scripting, regenerating audio, and editing. But it wasn’t fully authentic either. It sat in the middle, and the middle has no moat. Authenticity is the moat now.

AI TRAINING

🛍️ Give your AI agent a credit card (safely)

The Rundown: In this guide, you will learn how to let an AI agent make a real online purchase without handing over your real card. We tested Agentcard, which gives an AI agent a capped prepaid card, by having Codex buy a low-cost item online.

Disclaimer: If you try this, read the Agentcard docs, only connect through Agentcard’s official Stripe connection, and fund it with a prepaid or low-limit card.

Step-by-step:

  1. Install the AgentCard CLI, sign in, complete the setup through the official Stripe flow, then create a low-limit virtual card for a small test purchase

  2. Give your AI agent (we used Codex) a tightly scoped checkout task: one merchant, one product, a maximum budget, use the AgentCard virtual card for payment, and stop before clicking Place Order

  3. Watch as the agent works, approve any login, verification, or payment prompts, and only authorize the purchase after reviewing the final checkout screen

  4. After the purchase, review the transaction history and remaining balance, then close the virtual card so it can’t be used again

Pro tip: Try AgentCard’s built-in shopping tools for supported merchants like DoorDash, which let agents search, budget, check out, and reorder within limits.

PRESENTED BY IBM

💸 What do you fund when ROI isn’t clear?

The Rundown: AI isn’t about ROI yet — it’s a positioning move that reshapes how value is generated and how advantage accumulates. The CFO’s job isn’t defending AI spend; it’s making the cost of delay visible, especially with 55% of executives tying future advantage to execution speed.

Finance can set the pace — if it shifts:

  • From explaining variance to designing financial systems

  • From control to conditions for experimentation

  • From oversight to scaling AI with the C-suite

Read more in ‘Go big or get left behind.’

ANTHROPIC

🛡️ Anthropic flags Alibaba’s ‘largest’ distillation attack

Image source: Images 2.0 / The Rundown

The Rundown: Anthropic just accused Alibaba of running the largest known distillation attack, with 28.8M Claude exchanges extracted through nearly 25K fraudulent accounts in 45 days — targeting its AI models’ most advanced capabilities.

The details:

  • Distillation attacks work by querying a frontier model at scale and using the responses to improve a weaker one, capturing advanced capabilities.

  • In this attack, disclosed in a letter to the Senate Banking Committee, Claude’s agentic reasoning, coding, and long-horizon task capabilities were targeted.

  • In Feb., Anthropic had disclosed attacks by DeepSeek, Moonshot, and MiniMax involving 16M+ exchanges. OpenAI also raised similar concerns.

  • Anthropic is calling for antitrust clarity to let AI labs share threat intel, stronger chip export controls, and sanctions against Chinese labs behind distillation.

Why it matters: Distillation is a legitimate technique — every major lab does it. But there’s a difference between shrinking your own model and systematically harvesting a competitor’s capabilities. China’s AI progress has been real and impressive. The question is how much of it was earned, and how much was extracted.

QUICK HITS

🛠️ Trending AI Tools

  • ☁️ Render – Modern cloud platform for AI-native and multi-service apps. Ship faster, scale reliably, no infra complexity

  • 🤖 Gemini 3.5 Flash – Google’s ultra-fast model, now with computer use

  • 🧠 Ornith-1.0 – Self-improving AI coding model rivaling Opus 4.7

  • 🔀 Fusion – OpenRouter’s compound AI system with Fable-level intelligence

*Sponsored Listing

📰 Everything else in AI today

Meta Superintelligence Labs reportedly hired the founders and team behind AI security startup Virtue AI to strengthen its AI safety and agent security efforts.

Apple raised prices mid-cycle across several product categories, with Macs and iPads hit hardest, as memory and storage prices continue to surge in the age of AI.

Microsoft introduced AI Skills for Copilot in Excel, enabling reusable AI workflows, including pre-built skills for financial modeling, forecasting, and variance analysis.

U.S. giants, including OpenAI, Anthropic, Microsoft, and Amazon, backed the launch of RAISE US, a $500M initiative to help Americans prepare for AI job disruption.

Google is reportedly reorganizing its AI coding strike team into a dedicated “midtraining” group to catch up with Anthropic as key researchers leave for the rival lab.

COMMUNITY

🤝 Community AI workflows

Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.

Today’s workflow comes from reader Ly-ann in Singapore:

“I conduct my classes through Zoom. Since Zoom has the audio transcript feature, I save each one into a folder called ‘Class Transcripts.’ At the end of the week, Co-Work runs a scheduled task to review all the interactions in the classes I’ve had, highlights specific moments to flag, strengths, watch points, and next steps, against 6 frameworks I prioritize as an online educator. (CoI Facilitator, Online Learning Synthesist, Social Network Interaction Designer, and TBL Facilitator.)

It then suggests a few specific things I can do to improve my teaching. This way, I get incrementally better and better at my job!”

How do you use AI? Tell us here.

🎓 Highlights: News, Guides & Events

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown

Liked Liked