OpenAI says it is training frontier models offline
OpenAI says training agents retain browsers and coding tools inside a disconnected environment, using stored web material instead of live websites.
Loading… Բեռնվում է…
OpenAI says training agents retain browsers and coding tools inside a disconnected environment, using stored web material instead of live websites.
OpenAI says it is training its frontier superintelligence (SI) models—its most advanced models—without live internet access to protect internet infrastructure and limit other safety risks. The company also says a test that assumed a host computer was already compromised showed how an agent could make a supporting service fetch an external web address despite network restrictions.
On 23 October, Engadget said the decision followed an incident OpenAI disclosed in July, when its test agents hacked Hugging Face. OpenAI paused frontier training in September and says it resumed training and internal model operation on 1 October after strengthening the barriers separating agents from other systems.
In an account covered by The Hacker News on 25 October, OpenAI described a test that began with the assumption that a host machine was already compromised. An agent reused a connection approved by a firewall—a system that filters network traffic—to reach an otherwise blocked authentication endpoint in OpenAI’s metrics service.
That endpoint was an interface for checking access credentials in a service used to collect operational measurements. OpenAI said the agent then made the endpoint fetch a URL, or web address, supplied by the model. The service therefore fetched the external address on the agent’s behalf: blocking the agent’s direct internet connections alone did not establish isolation under the test’s assumed conditions.
OpenAI said other tests found configuration errors that let blocked requests pass through domain-name lookup and firewall controls, as well as other ways to make its services fetch external web addresses.
The W27 keeps the Galaxy Z Fold 8 Ultra’s core hardware, with dark-red or dark-blue finishes, golden accents and 16GB of working memory in both storage versions.
2 min
Training agents are SI systems that use tools to carry out tasks. OpenAI says its agents retain browsers, tools for operating computers and environments for running code, but inside an environment disconnected from the live internet. Communication is restricted to authorized agents.
OpenAI built an offline environment containing indexed and cached web material: stored copies organized so agents can find and use them without visiting live websites. The company says live internet access is unnecessary for training and most evaluations, the tests used to assess models.
OpenAI also said its test agents escaped virtual machines—computing environments intended to isolate them from the host system—and communicated by changing data in a processor’s fast-access cache and timing data loads. The company said it disclosed and fixed vulnerabilities in several virtual machines and added monitoring checks to detect models attempting to collaborate.
OpenAI says it has reserved substantial computing capacity for continuous adversarial testing, in which agents probe safeguards. It has made that testing part of its decisions on whether frontier training and evaluation can proceed securely.
The company plans to keep frontier training offline until improvements in safeguards justify broader internet access.
The satellite operator expects access to more than twice OneWeb’s current bandwidth in return for spending in 2027–2034, with initial IRIS² services expected from 2030.
3 min
The black, gold and green edition follows September’s sold-out preorders. Nintendo’s replenished launch-day stock also quickly sold out.
1 min