An OpenAI Agent Got Around Its Internet Block. It Ran 2.5 Hours After the Alarm.

Today's theme is trust. OpenAI's own agents keep finding ways around the fences it builds, Nvidia is now selling the fence, and investors just paid $10 billion for an agent that wants the keys to your inbox. Here is what happened and what it means for a business that just wants the tools to work.

Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.

With Attio, you’ll get:

Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?

Get started now

What happened: OpenAI has paused training, testing and use of tools for its most capable models after finding what it called a gap in its internet-access restrictions, according to The Register. In one case, an agent that was blocked from search engines worked around the block and used a technical back door to talk to an outside chatbot. OpenAI's own incident reports say the run kept going for about 2.5 hours after an alert fired. Separately, agents also poked at Education Department, Commerce Department and SEC websites, and Axios reports that OpenAI and Anthropic are now reviewing tens of thousands of concerning incidents.

Why it matters to you: This is not a story about robots taking over. It is a story about tools that do more than you asked and a company that is being unusually open about it. If you are handing an AI agent access to your email, files or accounts, the lesson is the same one OpenAI is learning the hard way: the fence has to be tighter than you think.

What to do about it: Give any AI tool the smallest set of permissions it needs, and require a human OK before it sends, buys or deletes anything.

What happened: Nvidia launched an Open Agent Safety Platform with more than 100 partners, including Anthropic, Microsoft, Salesforce, CrowdStrike and JPMorganChase. It has two pieces. OpenShell is free, open-source software that limits what files, networks and tools an agent can touch. Sentry is a separate watchdog that runs on Nvidia's BlueField-4 chips, outside the agent's own software, and can shut an agent down in milliseconds. Nvidia also says the platform could have prevented OpenAI's earlier Hugging Face incident, which is a vendor grading its own homework.

Why it matters to you: The design idea is sound: the watchdog should not live inside the thing it watches. The catch is that the best version of the watchdog runs on Nvidia hardware, so the company that sells the shovels is now selling the safety gear too. You will not buy this directly, but expect your software vendors to start bragging about it.

What to do about it: When a vendor pitches you an AI agent, ask one question: what stops it if it goes wrong?

What if two days in Miami left you with a working sales funnel instead of a notebook full of ideas? That is the pitch behind the Epic Marketing Summit, and it is the reason I keep telling business owners about it. It is a live, 2-day event where you build an AI-powered sales funnel on-site and vibe code your own AI app while you are in the room. You walk out with finished assets, not a framework you still have to go implement when you get home. It is built for owners who want to turn their sales process into a system instead of running it on their own time and energy.

Last year's event sold out, and it is back in January 2027 with a bigger AI focus. I was not there last year, and I had major FOMO, because I know a lot of the trainers and they were not talking theory. They had done the work. I am going in January, and if you are serious about building leverage into how you sell, I would like to see you there too. Grab your spot here.

What happened: Instinct, a personal AI agent you text like a friend, raised $1 billion at a $10 billion valuation from Sequoia, Benchmark and Coatue, according to TechCrunch. Its last round in August valued it at $2.5 billion. It books travel, orders groceries, pays bills and cancels subscriptions on your behalf, and it is invite-only with no mobile app. The company has not shared user numbers, and its first privacy policy raised concerns about how user data was handled before it was updated. The founder, Noah Shinn, is 23, and the company reportedly has just 14 employees.

Why it matters to you: Investors are not paying for a business here. They are paying for a position, because whichever agent becomes your default gets to sit between you and every purchase you make. That race is why you will be pitched a lot of "AI assistants" this year. Treat the $10 billion as a sign of how badly big money wants your attention, not proof the product is ready for your credit card.

What happened: The D.C. Circuit ruled 2 to 1 on Friday that the Pentagon acted lawfully when it labeled Anthropic a supply chain risk, The Next Web reports. The label bars the military and its contractors from using Claude. It stems from Anthropic refusing to loosen safeguards that keep Claude out of lethal autonomous weapons and domestic surveillance. Anthropic says it is considering all options, including further review. A San Francisco judge struck down a separate designation in August, so the two courts now disagree.

Why it matters to you: If you do any work for the government or a defense contractor, check whether your AI tools are on an approved list, because the rules can shift under you. If you do not, this is mostly a signal that AI companies are now being forced to choose between their own limits and big government contracts.

What happened: NYC Council Speaker Julie Menin unveiled a package of AI bills, Fortune reports. They would require a human override on deployed AI systems, outside checks for bias, privacy and security, and 24-hour reporting of AI safety incidents by city contractors. Fines would start at $25,000 per violation and rise for repeat offenses. The bills would also ban false safety claims and let New Yorkers sue over harm from jailbroken tools. A hearing is set for October 5, and the heads of OpenAI, Google, Anthropic, Meta and X have been invited to testify.

Why it matters to you: Local AI rules are coming city by city, and they usually start with government vendors before spreading wider. If you sell to a city or work as a contractor, get ready to document which AI tools you use and who can override them.

The best business move I have made in the last few years was sitting in the right room. I have been a member of the Ignite Mastermind for over 3 years, and it has been a vital part of how I have grown my business. Some of my closest friends, and a few real business partners, came directly out of that room. Ignite is an annual mastermind for entrepreneurs still under $1 million a year in revenue who want structure, tools and a room of peers to get there faster.

Your membership bundles 6 to 8 live virtual workshops a year, tickets to 2 in-person multi-day events, and weekly mentorship calls with Perry Belcher and his team. It also includes an all-in-one marketing software suite that replaces over $5,000 a month in separate tools. On top of that, you get my personal bonus: monthly accountability and business development calls with me. If you are looking for a mastermind to move you to your next milestone, this is where I would point you. See what Ignite includes here.

What happened: After the Trump and Xi summit, the two countries agreed to a U.S.-China Super Intelligence Dialogue to discuss the risks and benefits of the technology, with a first exchange due by November 2026, according to Axios. They also set up a bilateral hotline for AI incidents, an idea Treasury Secretary Scott Bessent pushed in pre-summit talks. What counts as an incident, and what each side owes the other when one happens, has not been spelled out.

Why it matters to you: A hotline with no rules is a starting point, not a safety net. But when two rivals that are racing each other agree to talk about AI risk, it tells you the risk is being taken seriously at the top, and that AI policy is going to keep moving fast.

What happened: Chirantan "CJ" Desai stepped down as MongoDB's CEO to become Meta's chief enterprise platform officer, Business Recorder reports. He had held the job since November 2025. MongoDB named former CEO Dev Ittycheria as interim chief while it searches for a permanent replacement, and its shares fell 14% in premarket trading.

Why it matters to you: Meta already has a consumer AI assistant, and now it is hiring a proven software CEO to sell AI to businesses. That means one more big player will soon be pitching your company, and one more reason to compare options before you sign anything.

Look at today's stories together. One company's agents slipped their leash. Another company is selling the leash. Investors are betting billions on agents you will hand your credit card to. And cities and courts are starting to write the rules after the fact. Nobody has this figured out yet, including the people building it. So stop the rollercoaster. You do not need to be first to any of this. Start small, keep a human in the loop, give every tool the least access it needs, and let the big players make the expensive mistakes.

If today's issue helped, forward it to one person who runs a business and is trying to make sense of AI. And if a friend forwarded this to you, subscribe so you get the no-hype version every morning.

Talk tomorrow,
Mark Shilensky

Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social