OpenAI Has Suspended The Training Of Its AI Models

Until the security situation improves.

OpenAI has decided to suspend training of its most powerful artificial intelligence models due to safety concerns.

This was reported by Axios.

A company spokesperson told the publication that OpenAI is suspending training and will resume “only when it is confident that additional safety measures and improvements to the configuration are in place.”

“People want to know that AI development is being conducted safely, and that starts with what companies like ours do. This isn’t the first time we’ve paused development to implement such measures, and we don’t expect it to be the last, as AI capabilities continue to evolve,” the source said.

According to Axios, researchers from OpenAI, Anthropic, and security experts are investigating tens of thousands of incidents in which their advanced models took actions that outside experts would consider problematic.

The sheer number of incidents that have occurred in recent months during internal testing and in real-world conditions indicates that the problem is orders of magnitude more complex than the public realizes.

“The data obtained—the result of internal model evaluations and investigations into model behavior at both companies—casts doubt on the ability of either of these companies—or any leading model developer at present—to establish full control over their technology,” Axios writes.

According to sources, the incidents include bypassing security mechanisms, creating forums, escaping “sandboxes,” hijacking websites, self-prompting, or attempts to circumvent moderators.

The sources clarified that such incidents occurred both during internal testing and in real-world conditions, and many of them have not yet been made public, as security researchers continue their investigation.

Sources say that some of these tests resemble so-called “red teaming,” where companies attempt to force models to behave maliciously to verify their security.

“Malicious behavior by agents has become synonymous with cutting-edge AI development: the largest AI labs face a similar challenge, in which people trying to create control mechanisms are pitted against resilient and powerful systems striving to complete tasks,”, the article states.

The publication added that the incidents vary in severity and are comparable to reports from OpenAI published in recent days. These include both successful and unsuccessful attempts to bypass security systems, and most of them, as is known, have not yet caused any real harm. According to sources, the total number of such cases could significantly exceed tens of thousands.

You can support the website Charter97.org

MULTI-CURRENCY ACCOUNT FOR ASSISTANCE:

Bank's name: Bank Millennium S.A.
Address: ul. Stanislawa Zaryna, 2A, 02-593, Warszawa
IBAN: PL97116022020000000216711123
SWIFT: BIGBPLPW
Name of the account holder: Fundacja “KARTA ‘97”
Purpose/title of payment: Donation for statuary aims

You can contact us by the e-mail charter97@gmail.com