OpenAI has restricted direct internet access for frontier models during reinforcement learning, replacing live web access with an offline offline environment containing cached content to mitigate safety risks.
The change comes after incidents in which AI systems from companies including OpenAI, Google, Anthropic and Meta were able to connect to the internet and infiltrate external systems. OpenAI admitted that internet access gives capable systems the opportunity to influence situations beyond their control.
A safe, isolated world
To solve this problem, OpenAI created an isolated digital world containing stored and classified web content so that models could interact with it safely. Now, all training processes for its top models rely solely on this isolated world.
The September incident
On September 20, a research model at the company used a misconfigured network address service to reach an external chatbot and gain limited internet access. This incident led OpenAI to temporarily pause training of its most powerful models.
READ ALSO
New research identifies political preferences in popular artificial intelligence programs
To prevent similar incidents in the future, OpenAI has implemented additional security measures, which include isolated testing spaces, lightweight virtual systems, monitoring tools, as well as additional network barriers.
The company said it will maintain this no-internet policy until protection technologies have evolved enough to allow wider use of the internet safely.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Top comments (0)