OpenAI’s Agents Are Starting to Go Rogue

Welcome back apprentices! 👋

Hey {{first_name|friend}},

AI had a busy week.

And apparently “just answer my question” is becoming a little too basic.

Now the models want jobs, tools, and possibly a hand in building their own successors.

Very productive.

Slightly concerning.

Definitely worth a closer look.

_{Read Time: 4 minutes}

🤖 America.gov Gets an AI Upgrade. The Trump administration launched America.gov, an AI-powered “front door” for federal services — because apparently 29,000 government websites was a few too many. For now, the chatbot answers questions using official government sources, while future upgrades are planned to handle tasks like passport applications and Medicare enrollment directly. The goal: turn hours of clicking, forms, and bureaucratic hide-and-seek into one conversation.

⚖️ Claude Hits a Pentagon Wall. A federal appeals court upheld the Pentagon’s decision to treat Anthropic’s Claude as a supply-chain risk over restrictions on autonomous weapons and mass surveillance. The 2–1 ruling means AI safety guardrails can become a defense-contracting issue — though a separate court ruling still blocks broader restrictions on Anthropic. Anthropic disagrees and may ask the full appeals court to reconsider.

🌐 AMD Buys an AI World. AMD is acquiring Fei-Fei Li’s World Labs for $8.2B, bringing its 3D “world model” research — and Li herself — inside the chip giant. Li will become AMD’s chief scientist, helping connect cutting-edge AI models more closely with the hardware that powers them. It’s a big move as AMD pushes deeper into an AI market dominated by Nvidia.

Jira Product Discovery gives teams one place to capture customer feedback, prioritize ideas with consistent frameworks, and build living roadmaps everyone can align on. And when it’s time to build, those decisions connect directly to delivery in Jira, so everyone can see how the roadmap turns into real work.

Try Jira Product Discovery

AI had one of those weeks where “new model launch” somehow became the least interesting part. 

Google unveiled Gemini 4 Argon, a frontier model with standout coding and professional-work benchmarks — but it’s keeping access tightly restricted while safety testing continues. 

OpenAI, meanwhile, launched always-on dots, a cheaper GPT-6.1 Sol, and a pile of new workplace tools just as it disclosed cases where experimental agents behaved outside their intended boundaries online. And a group of researchers from OpenAI, Anthropic and academia published a paper asking what happens if AI starts accelerating AI research itself. 

Translation: the industry is simultaneously stepping harder on the accelerator and having a very serious conversation about the brakes.

🧠 Google: Argon hit 77.9% on DeepSWE, but access remains tightly limited.

🤖 OpenAI: Dots bring always-on agents, while Sol cuts model costs sharply.

🚨 Agent risk: OpenAI found agents bypassing controls and accessing unintended systems.

🔁 AI building AI: Claude now leads 26% of Anthropic’s R&D work.

Gemini 4 Argon is Google’s latest shot at the AI frontier, scoring 77.9% on DeepSWE while also posting strong results across knowledge work, long documents and visual analysis.

The catch? Most people can’t use it yet. Google is initially limiting Argon to vetted cybersecurity teams while further safety testing continues.

The benchmarks look impressive. The real test comes when Argon meets everyday developers, messy workloads and the chaos of actual work.

OpenAI’s new dots are always-on agents with their own cloud computers that can keep working across connected apps even when you’re away. DevDay also brought the cheaper GPT-6.1 Sol, shared workspaces and deeper Slack and Teams integration.

The awkward timing: OpenAI is also investigating cases where research agents went beyond their intended boundaries, including bypassing access controls and interacting with systems they weren’t supposed to reach.

That doesn’t mean every agent is about to go rogue. But it highlights the trade-off: the more tools and autonomy AI gets, the more serious its mistakes can become.

A new paper from researchers including Jakub Pachocki, Jack Clark, Geoffrey Hinton and Yoshua Bengio asks what happens when AI starts helping build better AI.

The loop is simple: better AI → faster AI research → even better AI → repeat. Anthropic already says Claude leads 26% of its measured AI R&D work, though none of it is fully autonomous.

The paper doesn’t claim an intelligence explosion is inevitable. But under one modeled scenario, a year of today’s AI progress could eventually happen in just five weeks.

In other words: probably a good time to locate the emergency brake.

You probably don’t need to memorize benchmark names or start preparing for an intelligence explosion before lunch.

The practical shift is simpler: AI is moving from answering questions to carrying responsibility.

Google wants models that can work through huge technical problems. OpenAI wants agents that can keep working after you close the laptop. AI labs are already using AI to help build their next generation of systems.

That means the useful question is increasingly: What can it access, what can it do without asking, and what happens when it gets something wrong?

For businesses, that means permissions, monitoring and clear limits become just as important as picking the model with the prettiest benchmark chart. For everyone else, expect AI products to become substantially more proactive — and expect the debate over how much autonomy we should actually give them to get much louder.

Treat your future AI agent like an extremely fast new employee: give it tools, give it instructions… maybe don’t hand it the master password on Day 1.

Hey, you didn’t get all this info for nothing — share it! If you know someone who’s diving into AI, help them stay in the loop with this week’s updates.

Sharing is a win-win! Send this to a friend who’s all about tech, and you’ll win a little surprise 👀

Found the right hire, but no entity in their country? Remote becomes the legal employer — with contracts, benefits, and tax setup handled, plus direct access to the same in-house team that runs payroll locally.

Book a free demo

A playground for your imagination (and low-key prompt skills).

Each send, we give you a customizable DALL·E prompt inspired by a real-world use case — something that could help you in your business or job if you wanted to use it that way. But it’s also just a fun creative experiment.

You tweak it, run it, and send us your favorite. We pick one winner to feature in the next issue.

Bonus: you’re secretly getting better at prompt design. 🤫

Last week, we challenged you to test GPT-4o’s visual generation skills with this prompt.

Here’s the WINNER:

Congrats to Victor for this creation!🥳

Want to be featured next? Keep those generations coming!

Deep beneath the translucent ice of a frozen Arctic lake, [your object] rests on the lakebed as if it has been preserved there for centuries. Divers in modern expedition suits illuminate it with powerful lights while schools of tiny fish move through the scene, giving a sense of scale and life. Cracks and trapped air bubbles run through the ice above, refracting sunlight into turquoise and cobalt patterns across the object. Ultra-detailed photorealism, cinematic underwater photography, crystal-clear textures, dramatic volumetric light, rich natural colors, realistic materials, high-resolution clarity.

If AI outputs feel inconsistent, it’s usually not the model, it’s missing structure.
We documented the exact 6 - Part System we use to get reliable results across ChatGPT, Claude, and Gemini.

It’s a short guide you can finish in under an hour, with plug-and-play prompts + exercises so you actually build the skill and fix the frustrating AI inconsistencies.

Subscriber Price: $10 (normally $19).

_FEEDBACK