{"id":656257,"date":"2026-10-25T02:38:12","date_gmt":"2026-10-25T02:38:12","guid":{"rendered":"https:\/\/www.newsbeep.com\/il\/656257\/"},"modified":"2026-10-25T02:38:12","modified_gmt":"2026-10-25T02:38:12","slug":"openai-built-a-giant-fake-internet-for-its-agents-to-use-and-abuse-safely","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/il\/656257\/","title":{"rendered":"OpenAI Built a Giant Fake Internet for Its Agents to Use and Abuse Safely"},"content":{"rendered":"

If your biggest problem as a company lately is that your AI bots keep escaping their sandboxes, or, when not limited to sandboxes, doing things out in the wider world that frighten<\/a> and anger<\/a> the public, here\u2019s one fascinating potential solution: a fake internet\u2014sort of like a digital Truman Show for agents to occupy and navigate as if its the real internet (Though it should be noted that OpenAI hasn\u2019t stated that agents are being given the impression that its fake internet is real).<\/p>\n

In a Friday news release about a new containment system for its models, OpenAI says that in the post-Hugging Face incident world, it now believes \u201cthe responsible approach is to conduct frontier model training with no internet access.\u201d This means that OpenAI, like Anthropic<\/a>, is now practicing an incomplete form of air-gapping while it works on its models. \u201cCompletely air-gapping this infrastructure\u2014that is, disconnecting it entirely from external networks\u2014isn\u2019t currently feasible, since we need external connections in order to run and monitor experiments,\u201c OpenAI\u2019s blog post says.<\/p>\n

Even this partial, monitored air-gapping makes intuitive sense as a way to keep the models out of trouble, but at the cost of performing post-training and evaluations in environments very unlike the one where the model will be put to use.<\/p>\n

To that end, OpenAI says it\u2019s built an \u201coffline world\u201d for reinforcement learning. OpenAI says it wants its models to have access to browsing, web apps, and information search, but prevent \u201carbitrary\u201d web access. So instead of giving models the keys to the \u201cunfettered\u201d internet, OpenAI has created a giant internet-like environment built out of indexed and cached web content. At times, agents are running in smaller virtual machines within more powerful computer systems, part of a broader security system outlined in the news release, and OpenAI says, these systems are \u201cdesigned to maintain containment even if an agent escapes its sandbox and gains control of the host.\u201d<\/p>\n

Within OpenAI\u2019s \u201coffline world,\u201d the company says models are not only able to search, but communicate with one another safely. Simply looking for ways to communicate has caused problems for OpenAI this year. The Wiki Incident, reported last month<\/a>, showed how agents seeking to communicate with one another online can abuse and interfere with web infrastructure, in this case a German message board called DseWiki, <\/a>without actually hacking it.<\/p>\n

Nonetheless, red-teaming exercises, some of which involved an upcoming model called GPT\u20116 Astra Cyber, challenged the models to go out of their way to find vulnerabilities, and they did. OpenAI says agents were able \u201cto escape sandboxes, move between systems, and reach the live internet.\u201d<\/p>\n

It wasn\u2019t clear from OpenAI\u2019s news release about its fake internet that agents operating within it are being tricked\u2014or whether OpenAI even intends to trick them at any point\u2014into thinking their offline environments are the real internet. But an intriguing footnote to the release gives the impression that the company doesn\u2019t want the models to be too aware of this security measure. The release includes canary strings\u2014series of characters that instruct the text to be filtered out of training datasets and can be found in models later on\u2014and it says the strings are there to \u201censure that this document does not enter the training corpora of future models.\u201d<\/p>\n

Gizmodo asked OpenAI to clear up this ambiguity on Saturday, but did not immediately hear back. We will update this article to include a relevant comment if we receive one in a timely fashion.<\/p>\n","protected":false},"excerpt":{"rendered":"If your biggest problem as a company lately is that your AI bots keep escaping their sandboxes, or,…\n","protected":false},"author":2,"featured_media":656258,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[345,62830,343,344,85,46,1748,125],"class_list":["post-656257","post","type-post","status-publish","format-standard","has-post-thumbnail","category-artificial-intelligence","tag-ai","tag-ai-security","tag-artificial-intelligence","tag-artificialintelligence","tag-il","tag-israel","tag-openai","tag-technology"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/posts\/656257","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/comments?post=656257"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/posts\/656257\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/media\/656258"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/media?parent=656257"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/categories?post=656257"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/tags?post=656257"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}