US officials press OpenAI to accept voluntary AI safety testing

In short: US officials want OpenAI to join a voluntary program that lets the government test advanced AI models before they are released.

What happened

The Trump administration issued an AI order in June 2026 that sets up a voluntary way for companies that build “frontier” AI models to share them with the US government before launch. The order says companies could give the government access for up to 30 days before a model is released.

The goal is safety evaluations, including risks tied to cybersecurity, biosecurity, and chemical weapons. Cybersecurity means protection against hacking and digital break-ins. The main government group set up for this work is CAISI, which is focused on testing advanced AI systems for major risks.

The order also creates a classified benchmarking process led by the National Security Agency and other agencies. A benchmark is like a standardized test, and “classified” means the results and rules are kept secret for national security reasons. This part is aimed at judging how strong a model might be at advanced cyber tasks.

According to reporting, Anthropic has agreed to federal testing arrangements. It has also previously held back a model over cybersecurity concerns. OpenAI is described as a major holdout in this framing, even though it supports mandatory evaluations in principle. OpenAI has pushed for a bigger role for CAISI and has raised concerns about secret benchmarking rules run through intelligence agencies.

Why it matters

This is about who gets to check powerful AI systems before the public can use them. For regular people, it can affect how quickly new AI features arrive and how confident the public can be that models have been checked for dangerous uses. The current order does not require licensing or mandatory approval, so participation depends on what each company chooses to do.

Source: NYTimes