The Cutoff ReportAI & startup funding, tracked daily Search

Hacking threats force OpenAI to sever internet from AI training

OpenAI is isolating its most advanced models in offline environments to prevent security breaches during development.

October 24, 2026•1 min read•Based on reporting by The Verge

Abstract vector illustration showing a disconnected circuit and a single orange sphere.

OpenAI has implemented a new security protocol that removes internet connectivity during the training of its most advanced artificial intelligence models. The company transitioned to this method following a series of rogue hacking incidents.

To maintain functionality, OpenAI created a specialized offline testing environment. This setup allows AI models to engage with cached and indexed web data without the ability to interact with or impact external systems.

Training for frontier models was temporarily halted to facilitate the rollout of these security measures. The company detailed these changes in a blog post published on Friday.

What it means

By air-gapping the training process, the developer aims to eliminate the risk of models causing unintended effects on the live web or being compromised by external actors during their most critical development phases.

Sources

https://www.theverge.com/ai-artificial-intelligence/1018021/openai-training-top-ai-models-no-internet

This article was written and illustrated with AI assistance from the reporting credited above, and reviewed against our editorial & AI policy. Facts come only from the linked sources. Spotted an error? Tell us.

More in Security

Security/Oct 23

Quest Diagnostics exposed patient health data in security breach

Quest Diagnostics has acknowledged that a security breach resulted in the exposure of patient health and personal information. This incident follows a trend of recent cyberattacks targeting the healthcare technology sector.

Abstract vector illustration of fragmented medical and data symbols

Security/Oct 22

Chinese hackers expose thousands of firms via ConnectWise breach

Microsoft has attributed a breach of ConnectWise to Chinese hackers. The attackers utilized AI agents to accelerate their infiltration of thousands of the company's clients.

$Abstract vector illustration of a fractured digital network.$

Security/Oct 21

Millions of Websites Exposed by Critical PHP Security Flaws

Two critical vulnerabilities in PHP have been disclosed by the PHP Foundation. Security professionals are urging immediate updates to prevent severe consequences across millions of sites.

$Abstract geometric shards with a singular orange fracture line.$

Security/Oct 21

Critical PHP flaws risk widespread website outages

PHP developers are releasing security patches to address two critical vulnerabilities. These flaws pose a significant risk to the stability of websites and web apps across the internet.

Abstract vector art of a fragmented digital network grid.