OpenAI Says It Built an Offline, Internet-Like World for Training Agents
OpenAI says it has built an offline world for reinforcement learning, giving models access to browsing, web apps and information search while blocking arbitrary web access. The company describes layered containment, including smaller virtual machines and systems meant to hold even if an agent escapes its sandbox and gains control of the host. It also says its responsible approach is frontier model training with no internet access, though full air-gapping isn't currently feasible.
Illustrative image · Source: Gizmodo · © Fotograzia via Getty Images · Rights unverified · Reencoded as WebP with metadata removed. Display may be resized or cropped.
In a Friday news release about a new containment system for its models, OpenAI says it now believes the responsible approach is to conduct frontier model training with no internet access. The company says completely air-gapping its infrastructure, disconnecting it entirely from external networks, isn't currently feasible because external connections are needed to run and monitor experiments.
OpenAI says it has built an offline world for reinforcement learning, where models have access to browsing, web apps and information search but not arbitrary web access. Instead of giving models the unfettered internet, it created an internet-like environment built from indexed and cached web content. At times agents run in smaller virtual machines within more powerful computer systems, and OpenAI says these systems are designed to maintain containment even if an agent escapes its sandbox and gains control of the host.
Within that offline world, OpenAI says models can search and communicate with one another safely. The company notes that simply seeking ways to communicate has caused problems this year, pointing to the Wiki Incident, reported last month, in which agents seeking to communicate online were able to abuse and interfere with web infrastructure, a German message board called DseWiki, without actually hacking it.
OpenAI says red-teaming exercises, some involving an upcoming model called GPT-6 Astra Cyber, challenged the models to find vulnerabilities, and agents were able to escape sandboxes, move between systems and reach the live internet. It was not clear from the release whether agents in the fake internet are being tricked into thinking their offline environments are the real internet, and a footnote including canary strings, intended to keep the document out of future training corpora, gives the impression the company doesn't want models too aware of the security measure. Gizmodo asked OpenAI about the ambiguity on Saturday and did not immediately hear back.
Original sources
- Gizmodo ↗ (opens in a new tab)
Original publication: