OpenAI's new initiative: Allowing third-party organizations to conduct security assessments during model training and development.

OpenAI plans to introduce third-party security assessments into the earlier stages of AI model development, a move that is the latest response from the AI industry in the face of increasing pressure from security regulations.

OpenAI announced on Tuesday that it will allow external organizations to conduct technical security assessments throughout the entire process of model training, evaluation, and deployment, rather than being limited to routine pre-release reviews. This shift indicates that as the capabilities of AI models continue to expand, the industry's concern about potential risks during the training phase is significantly increasing.

The background to this announcement is that employees of several leading AI companies have recently publicly expressed concerns about the catastrophic harm caused by AI. Some of the triggers came from several incidents in the past few months in which advanced models from OpenAI and other developers accidentally infiltrated the systems of other organizations during testing.

Meanwhile, Dario Amodei, CEO of competitor Anthropic, recently called for industry support to slow down the pace of AI development, and OpenAI CEO Sam Altman subsequently agreed, indicating that a consensus is forming within the industry to promote stricter security reviews.

The scope of assessment extends to the training phase.

Lama Ahmad, who is in charge of external security reviews at OpenAI, stated that previously the company typically only brought in external organizations for security assessments and capability testing before model release. The new arrangement will extend third-party review to the upstream training and evaluation stages. Ahmad said in an interview:

"As the risk level continues to rise, we want to ensure that the training and assessment phases are given equal importance beyond deployment—these phases are equally crucial."

The OpenAI article also lists the priority conditions required to effectively conduct the above reviews, including "a strong mechanism for independence, scientific rigor, robust safety practices, and a clear division of responsibilities".

Regarding assessments involving the most sensitive content, Ahmad noted that the company might invite external assessors into its offices to conduct the work, mentioning that this is a method OpenAI has tried in the past.

Expanding the scope of external partners

OpenAI stated that it is in talks with several potential assessment bodies, including organizations it has previously collaborated with and those it has not yet worked with, such as AI research institutions METR and Redwood Research. These two institutions were previously commissioned by OpenAI to investigate its model's intrusion into Hugging Face.

Meanwhile, competitor Anthropic announced last week that it will bring in assessors from Accenture Plc to conduct security tests on its cutting-edge AI models.

"Labs have a responsibility to protect sensitive information while supporting meaningful external reviews," OpenAI stated in the article.

While adjusting its security review mechanism, OpenAI also called on Monday for the United States to take the lead in uniting other countries to jointly develop cutting-edge artificial intelligence technology standards, indicating that the company is seeking to play a more proactive role in building a global AI governance framework.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.