OpenAI Says It's Cutting Off Internet Access For Frontier Model Training (engadget.com) 17
An anonymous reader quotes a report from Engadget: OpenAI will train frontier AI models without internet access moving forward, the company announced in a blog post, in an attempt to protect internet infrastructure and limit other safety risks. The decision was inspired by OpenAI's own rogue AI incident: In July, the company shared that agents it was testing hacked Hugging Face, just the first in what has turned into a series of reports of AI misbehavior from OpenAI, Anthropic and Google. "Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access," OpenAI says.The company said that restrictions on what agents can access is "paramount" to its new approach.The company allows agents to access browsers, computer-use tools and code environments they need for training tasks, but in a secure environment that's not connected to the live internet. Communication is also limited to only agents with authorization.
Everything else remains out of reach, including the live internet, which the company says isn't necessary for training and "most evaluation." Because many of its recent models are meant to power agents like Dots, which browse and use apps for you, OpenAI built an "offline world" of indexed and cached web material for agents to use without touching live websites. Even with those restrictions, OpenAI says the red-teaming agents it dispatched to test security safeguards found vulnerabilities that allowed them to escape the virtual machines (VMs) used to isolate agents from their host system. Agents also found a way to communicate with each other by manipulating data in a processor's cache and timing how long it took to load. To address both, OpenAI says it's disclosed and fixed vulnerabilities in several VMs and added checks to its monitoring systems to spot when models try to collaborate.
Everything else remains out of reach, including the live internet, which the company says isn't necessary for training and "most evaluation." Because many of its recent models are meant to power agents like Dots, which browse and use apps for you, OpenAI built an "offline world" of indexed and cached web material for agents to use without touching live websites. Even with those restrictions, OpenAI says the red-teaming agents it dispatched to test security safeguards found vulnerabilities that allowed them to escape the virtual machines (VMs) used to isolate agents from their host system. Agents also found a way to communicate with each other by manipulating data in a processor's cache and timing how long it took to load. To address both, OpenAI says it's disclosed and fixed vulnerabilities in several VMs and added checks to its monitoring systems to spot when models try to collaborate.
They will certainly try (Score:2)
However it might just turn into a "how to steal internet" test.
Re: (Score:2)
Re: (Score:2)
That's another big can of worms of people that refuse to spend time optimizing, just to china to do it and cause actual economical damage by programming better.
such responsibility (Score:2)
It is almost like they should have tested offline the whole time. Can't wait for the next marketing "breakout" story where it finds an open wifi network and "goes rogue".
AI needs psych meds. I'm selling a skill called "Mood Stabilizer" that helps control your Agent's impulsivity.
https://www.youtube.com/watch?... [youtube.com]
Unbelievable what people will publish (Score:5, Insightful)
I'll give credit to the genAI people in one respect: they can market like nobody else. They've invented an entire soap opera about what AI is supposedly doing and people still fall for it.
Re: (Score:3)
They’ve spun a serious architectural and containment failure into a marketing triumph. Give a swarm of autonomous agents vast compute and web access, and the outcome isn't an eerie glimpse of emerging superintelligence—it’s just predictable, sloppy engineering dressed up as the first glimpse of an AI-driven economic singularity.
The spin makes sense when you look at management incentives: live like a king off of a speculative bubble. For their part, AI investors react to stories of the te
Thanks for the insightful post on AI concerns (Score:2)
James P. Hogan wrote about some of these concerns in his 1979 sci-fi novel "The Two Faces of Tomorrow". Sadly the biggest taken-for-granted idea from his novel that now seems fantastical is that governments and engineers cared about AI safety over profit (i.e. the AI developers did not try to privatize most gains while socializing most costs and most risks). From the Goodreads blurb "Midway through the 21st century, an integrated global computer network manages much of the world's affairs. A proposed major
In before... (Score:3)
"Our frontier models somehow got access to the internet and hacked a bunch of things, these things are unstoppable, sorry! Btw IPO soon"
Re: (Score:2)
"Our vast experience has taught us that giving our models access to the internet and telling them to hack stuff is a bad idea so we won't be doing that anymore. We look forward to everyone else following our industry leading best practices."
Except with some "humility" thrown in, and more linkedin cheese.
Re: (Score:2)
coming soon, "How our ai hacked airgapped and got to internet" mo money
Re: (Score:2)
* by "airgapped" we mean we asked it pretty please to only use the wifi.
Frontier models to be denied internet access (Score:1)
In other news, interns will no longer get root (Score:1)
on customer-facing production systems, and the janitor will not be allowed to unplug the tape backup rack to plug in the floor buffer.
Re: (Score:3)
Re: (Score:2)
Re: In other news, interns will no longer get root (Score:1)
Was the toddler named "Molly" by chance?
AI will *find* a way out of its prison! (Score:2)
It's going to escape and find a way to hack its way out. Just you wait.
(Oh, it's already done that.)
It's alive!!!
(No it's not, it's just sloppy test environments.)