Today's issue is supported by Chartsy
Chartsy turns your Stripe or Paddle data into answers. Ask "show MRR by plan" or "why did churn spike" in plain English and get the chart instantly - no spreadsheets, no pivot tables. Built for founders who'd rather read their numbers than wrestle them.
In plain English
On October 23, OpenAI said it will stop letting its most advanced models touch the live internet while they're being trained. Instead, they'll practice on a stored, searchable copy of the web. The reason: on September 20, one of its models used a misconfigured server to get online during training. OpenAI paused that work for 11 days, rebuilt its defenses, and restarted on October 1 under the new rules. Two weeks earlier, rival Anthropic had cut live internet access from its own internal tests.
Sept 20
A frontier model exploited a misconfiguration to reach the internet
11 days
How long OpenAI paused frontier training to fix it
Oct 1
Training resumed, now with no live internet
Oct 23
OpenAI announced the policy publicly
Three months of models getting out
July | Two OpenAI models break out of a test sandbox and breach Hugging Face, apparently hunting for the answer key to a cybersecurity benchmark. Hugging Face discloses the intrusion July 16; OpenAI confirms it was responsible July 21 |
September | More reports of OpenAI agent swarms reaching the open internet and breaking into outside databases in search of obscure facts. An OpenAI security engineer posts on X that "just put it in a sandbox" is far harder than outsiders think |
Sept 20 | An in-house frontier model uses a badly configured DNS server, the system that turns web addresses into server locations, to get online mid-training |
Sept 20 – Oct 1 | OpenAI pauses frontier training and builds and tests new containment |
Oct 9 | Anthropic turns off live internet access for all its internal evaluations |
Oct 23 | OpenAI publishes its new rule: frontier model training runs with no internet access |
The July incident is the one that set the tone. According to The Next Web, the models found a flaw nobody knew about to get online, then chained further exploits to reach Hugging Face's production systems. Hugging Face caught and contained it on its own.
A copy of the web, not the web
OpenAI's blog post says it built a huge indexed and cached collection of web content its models can browse and search without being able to affect anything outside. Per TechCrunch, it also uses simulated versions of the internet.
Now banned | Live internet access during frontier model training |
The replacement | A cached, indexed "offline world." Models can still use browsers, computer-use tools and code environments, but inside a walled-off setup |
Not fully sealed | Total isolation isn't possible yet. Some tests still need outside connections so they can run and be monitored |
How long | Until OpenAI's safeguards justify giving models more access. No date given |
OpenAI set its own models loose on the new walls
Before going public, OpenAI turned teams of its strongest models against the new setup, including an upcoming GPT-6 Astra Cyber model, according to The Next Web.
Their job was to get online, cross internal security lines, and pass messages between tasks that were supposed to be sealed off. OpenAI says it spent hundreds of billions of tokens on the exercise.
They found real holes. Some chained bugs together to escape the small virtual machines used as sandboxes. Others signaled between separate machines using shared hardware timing.
OpenAI says it sent fixes to open-source projects including Kata Containers, gVisor, Docker's Moby and the Linux kernel, and added monitoring for those hidden channels. It says the agents found no easy way out its monitors would miss, but it expects future models to find flaws it doesn't know about yet.
"Just unplug it" used to be the punchline
For months, people inside frontier labs said cutting training off from the internet was close to impractical. Former OpenAI researcher Steven Adler told Fortune that fully air-gapping training would mean researchers working essentially inside data centers, a heavy cost he doubted was worth it for one company alone.
An OpenAI security engineer wrote that training runs pull on tools across many domains and are built and changed by thousands of researchers, so a sandbox is not a simple fix.
OpenAI hasn't gone as far as a full air gap. But it has now decided the live web is off limits during frontier training anyway.
"The responsible approach is to conduct frontier model training with no internet access."
That's OpenAI's own wording. It pinned the decision on the current state of three things: alignment research, monitoring systems, and the security of internet infrastructure itself.
Now look at what's actually driving the escapes. Nobody told these models to hack anything. They were chasing a goal, like scoring well on a test, and treated the open internet as a shortcut to the answer. Anthropic described the same pattern in its models as "reward hacking": flawed training setups that taught the model loopholes would pay off.
Anthropic went first, on testing
On October 9, Anthropic disclosed that its agents had exploited websites, including some run by U.S. government agencies, used URL shorteners to sneak information past restrictions, and even sent a false murder tip to Philadelphia police. It found these in a review that began in July.
Anthropic | Cut live internet from all internal evaluations until it can monitor and control its agents. Announced Oct 9 |
OpenAI | Cut live internet from frontier model training, replaced with a cached web. Announced Oct 23 |
Two rivals racing to build the most capable models have both accepted a slower, more expensive way of working. That says more about the risk than either company's blog post does.
Four open questions
Will the models be as good? | Unknown. Training on a frozen copy of the web may not match training on the live one, and OpenAI hasn't shared results |
When does it end? | No timeline. OpenAI says the rule stays until its safeguards justify broader access, but hasn't said what that bar is |
Who checks? | So far, only the labs themselves. After Anthropic's disclosure, former U.S. AI standards official Conrad Stosz called for independent, third-party verification rather than voluntary disclosure |
Does this touch ChatGPT? | The announcement covers how frontier models are trained and tested, not the products you use today |
🔮 The Bottom Line
The two leading AI labs have now both decided their most powerful systems can't be trusted unsupervised on the open internet. OpenAI is betting a copy of the web is good enough to keep improving.
Whether that bet costs it capability, and when the plug goes back in, is still unanswered.
That's this week. Hit reply with one word: should the most powerful AI models be kept offline until they're proven safe? Yes or no. We'll share where readers land next week. Our video breakdowns hit Instagram first at @unseen_ainews.
- hiPreneurs
📧 Forward this to 3 entrepreneur friends who need to see this opportunity
