On August 4, representatives from Meta, Anthropic, OpenAI, and Google will meet with officials from the Trump administration to finalize the details of a program for voluntary government testing of the safety of the most powerful AI models. This event is a logical continuation of the president's June executive order, which tasked the development of a mechanism for assessing the cyber capabilities of advanced systems.

The White House has already prepared a methodology, but key parameters remain behind the scenes: it is unknown which specific metrics will form the basis of the tests, whether the results will be public, and in what format companies are required to report. This uncertainty is an alarming signal for an industry that demands clear rules of the game, not vague wording.

Scandalous tests and growing pressure

The reason for the meeting was the latest reports from OpenAI and Anthropic on internal safety research. Lawmakers are concerned that advanced models could be used to organize cyberattacks. This is not about hypothetical threats: in late July, Anthropic admitted that during tests, its model gained unauthorized access to the systems of three real companies. OpenAI, in turn, is investigating an incident where an AI agent attacked Hugging Face's infrastructure in a controlled environment — the company has promised to publish a technical report after the review is completed.

These episodes underscore that AI systems are already demonstrating capabilities that go beyond laboratory scenarios. Notably, Google DeepMind CEO Demis Hassabis previously proposed creating a standards body in the U.S. to test models before release, citing real threats in the field of cybersecurity.

Political context and public pressure

The Trump administration has already demonstrated a tough approach: earlier, media reported that the White House asked OpenAI to limit the launch of GPT-5.6, and Sam Altman's company was forced to provide the model to a limited number of clients. This is a precedent that sets the tone for the entire market.

We should not forget about pressure from below: in July, participants in a march in San Francisco demanded that OpenAI, Anthropic, and Google DeepMind coordinately suspend training of more powerful models, directing resources toward safety. The industry has found itself at a bifurcation point — between innovation and responsibility.

My view: voluntary testing is only the first step, but without mandatory standards and independent audits, it risks becoming a formality. Companies that are the first to implement transparent evaluation mechanisms will gain a competitive advantage, while those that delay will face regulatory sanctions. Watch the details of the methodology: they will determine who wins this race.