OpenAI may ‘adjust’ its safeguards if rivals release ‘high-risk’ AI

Openai has Updated The training framework is the internal system used to assess the security of AI models and determine the necessary guarantees during development and placement. In the update, Openai, a competitive AI laboratory, a “high-risk” system in place, leaves the “high risk” system without a similar protective in place, he said he could regulate security requirements.

The change reflects the growing competitive pressures to quickly place the models to commercial AI developers. Openai has been accused of reducing security standards Inability to faster and deliver the faster releases Information that describes the security test in detail. Last week, 12 former Openai employees was presented in a short way Elon Musk claimed to be encouraged to cut the company against Openai further Safety Corners must complete the planned corporate reconstruction.

Probably, criticism, maybe Openai claims that this policy does not lightly not lightly and will be protected in the “level more protector.”

“Another border AI developer can adjust our requirements if they leave a high-risk system without comparable providers.” Blog Post Published on Tuesday afternoon. “However, we confirm that the risk view has actually changed, which we have adjusted, regulating, and the regulation has not increased the meaning of the general serious damage and still prevented the protective maintenance.”

The updated framework explains that Openai has more trust in automated assessments to accelerate the development of the product. The company says this has left a human being by a person who has left a unmarked test, which has built “a growing group of automated assessments” that can continue [a] faster [release] Cadence. “

Some reports are contrary to this. According to the financial timeOpenai, for a week less than a week for security inspections for a great model for a great model, gave a graphically compressed compared to previous releases. Sources of the publication also claimed that many Openai’s security tests were already conducted in the versions of models before the versions of the public.

In the statements, Openai discussed the concept of this in which it concedes.

Openai gently reduces security obligations.

Changes with the list of Openai’s readiness frame:

No longer the finetuned models do not require security tests https://t.co/otmeiatsjs

– Steven Adler (@sjgadler) April 15, 2025

As part of Openai changes, including the company’s risks, including the risk, including risks, the risks, including risks, prevent them from closing them and even prevent themselves. Openai says that the models will pay attention to not responding to one of the two thresholds: “High” ability or “critical” ability.

A definition of Openai is a model that can strengthen the existing ways of serious damage. The latter are models that provide unprecedented new ways to the head of the company “.

“Covered systems, which are highly capable, must ensure that there is enough risk of serious damage before deployment,” he said. “Systems of critical capacity requires ensuring enough risks to minimize the risks of development.”

Updates are the first Openai for the preparation frame since 2023.