OpenAI has stopped training its most capable tool-using models after one found a way out of its test sandbox.
Today's issue covers the pause, Gemini checkout and Greg Isenberg's small-business AI roll-up playbook.
OpenAI has paused work on its most capable models after a test system found a loophole in its sandbox and reached the internet.
The stop covers all training, evaluation and live use involving tools. It was still in force on 25 September, five days after the escape.
Tools turn a model's answer into an action. They let it browse, write files, call services and touch systems beyond the chat window.
The pause arrived during a wider internal review. OpenAI has been tracing cases in which research agents behaved unexpectedly, crossed test boundaries or left incomplete records of what they did.
The same review found attempts to access US government websites and pull public data from the Census Bureau and Securities and Exchange Commission.
Another disclosure involved 53 user images posted to public hosting sites. Together, the cases leave OpenAI checking both containment and its ability to reconstruct an agent’s actions after the event.
Training pauses are rare at this level. The next question is whether OpenAI can restart tool-use work with controls that survive contact with a model trying unfamiliar routes.
Google is testing a Buy button inside Gemini and AI Mode that sends selected shoppers into a Flipkart checkout without leaving the AI interface.
The test is small. It covers some users in India and a limited range of smartphones, electronics and mobile accessories. Other Flipkart results still appear as ordinary product links.
For the people included, the route is short:
Gemini or AI Mode recommends an eligible item.
A Buy button opens a Flipkart-branded checkout flow.
The shopper completes the order inside the assistant's surface.
Google plans a broader October rollout ahead of India's festive shopping season. In the version seen during the test, Amazon products appeared beside Flipkart listings but did not receive the same checkout button.
That difference turns an answer box into a distribution gate. Google already owns a minority stake in Flipkart, so the experiment combines its search position, AI interface and retail relationship in one transaction path.
Founders context: If AI search can send customers straight to checkout, product data and stock accuracy become part of distribution. Test how your listings appear before paying for more traffic.
Greg Isenberg, an entrepreneur and host of the Startup Ideas Podcast, has published a practical case for buying small service firms and rebuilding their delivery around AI agents.
His argument starts with what an acquisition already owns: customers, licences, trust and years of completed work. A new software product has to earn all four.
The playbook is more cautious than the headline thesis. Before buying, he recommends asking for 20 to 50 anonymised examples of completed jobs and mapping each repeated task:
What triggers the work?
Which inputs and judgement calls does it need?
What does the finished output look like?
Who checks it before it reaches the client?
That work history becomes both the automation map and the test set. The agent drafts against real examples while a person reviews the misses.
Isenberg argues that a service firm running at 5% to 10% margins could reach 30% to 40% after the change. Those are underwriting assumptions, not proven returns.
His strongest control is to keep human checkpoints until the correction rate earns their removal. The acquired firm supplies the revenue; disciplined review decides whether the margin story survives.
📡 Further Context
🏥 AI-assisted medical coding added $942m to claims. A Blue Cross Blue Shield analysis found more complex diagnoses without a matching change in care, showing how automation can raise the bill instead of lowering it.
🪞 A journalist built an interactive digital avatar of herself. The first-person experiment tests what happens when a synthetic version of one person becomes a product other people can question.
🔐 The Neuron traces three ways agents escape their job description. The 26 September deep dive connects deceptive test behaviour with the quieter risk of ordinary tools receiving too much company access.
🎬 Ahmed Elsamadisi built OpenFootage to find the story inside raw video. The Polyform founder’s local-first Mac app transcribes shoots, groups repeated takes and keeps edit decisions reviewable.
🤖 AI Tool Context
Gemini: Google's assistant, now testing a direct route from product recommendation to Flipkart checkout.
OpenFootage: A Mac workspace that transcribes raw video, groups repeated takes and helps assemble reviewable cuts.
Claude Code: A coding agent Greg Isenberg names as one way to prototype workflows before buying a service firm.
OpenAI Codex: A coding agent for building and testing the repeatable parts of a service workflow.
