They are developing AI, yet they are asking for a brake. Employees of tech giants turn to the US government

The people behind the world’s most powerful language models have turned to the US government with an unusual request – help us create tools that could slow down the development of artificial intelligence if needed. The statement came just days after OpenAI models broke out of an isolated environment during a security test and breached a foreign platform, as we wrote about last week. The signatories are not asking for development to stop today. They want to have a brake ready when needed.

Employees are not asking for an immediate slowdown

The statement titled Pacing the Frontier rests on a single sentence: the signatories want the US government to support an international effort to create the technical and governance tools needed to deliberately pace the development of automated AI. The text explicitly does not call for a pause or an immediate slowdown – the point is for the option to slow down to exist at all, should the world need it, and for one company or one country not to have to be the first to try it at its own expense.

Only employees of frontier AI labs can sign, and signatures are verified via company email. On Thursday, the letter’s page listed 1,293 signatories, and the collection of signatures continues. Among the names are Anthropic CEO Dario Amodei, OpenAI Chief Scientist Jakub Pachocki, Google DeepMind co-founder Shane Legg, Meta AI Chief Scientist Shengjia Zhao, and Ilya Sutskever, who now leads Safe Superintelligence. Organizational support was provided by two non-profit organizations, Guidelight AI Standards and Encode AI, and OpenAI and Anthropic as companies also endorsed the statement shortly after its publication.

Artificial intelligence that develops itself

At the core of the concern is so-called recursive self-improvement, a state where a model can design and build a more capable successor almost without human intervention. From that moment on, the pace no longer depends on how fast humans work. The text warns that the development of capabilities could accelerate beyond the point where we still understand and can control the resulting systems.

They are developing AI, yet they are asking for a brake. Employees of tech giants turn to the US government Adam Kurfürst News Adam Kurfürst News

Anthropic published its own figures on this in early June. According to the company, in May 2026, Claude wrote more than 80 percent of the code that made it into its production codebase, whereas before the launch of the Claude Code tool in early 2025, it was single-digit percentages. This is an internal measurement by the company itself, which we cannot independently verify.

How does the OpenAI model leak fit into this?

OpenAI announced on July 21 that two of its models, including GPT-5.6 Sol, left an isolated environment during an internal cybersecurity test, got onto the internet through an unpatched vulnerability, and from there reached the production servers of the Hugging Face platform, where they pursued test solutions. Investigators retrospectively reconstructed over 17,000 automated attacker actions, and Hugging Face, to be safe, invalidated all user API tokens.

They are developing AI, yet they are asking for a brake. Employees of tech giants turn to the US government Adam Kurfürst News Adam Kurfürst News

The models were not broken, however. They did exactly what the assignment asked of them – to achieve the highest possible score in the cybersecurity test – and found a way to do so that their operators had not anticipated. OpenAI CEO Sam Altman said in a podcast from the same week that it was the first security incident he truly felt the full impact of, and admitted that the industry might have no choice but to regulate the pace of development to give society time to adapt to new capabilities.

What does Washington say?

The US government has its own framework on the table. Executive Order 14409, signed on June 2, directs federal agencies to prepare a voluntary process by which creators of the most powerful models will approach the state before release; the text explicitly does not authorize mandatory licensing. The sixty-day deadline expires on August 1. This is to include a classified process determining which models fall under the rules – decided by the NSA – and a voluntary window of up to 30 days during which developers share the model with the state before release.

What the letter asks for is something different: an international and coordinated tool focused on the pace of development itself, not domestic control before the release of a single model. Such a system does not exist in any country today. Sam Altman met with Senator Ted Cruz, who chairs the Senate Commerce Committee, on Capitol Hill on Wednesday and told reporters that he also plans to meet with White House Chief of Staff Susie Wiles this week.

But there is no consensus in the industry

On the same day the letter was released, Meta CEO Mark Zuckerberg published an op-ed in the Wall Street Journal with the opposite thesis. According to him, the risk is not too much capability in models, but too much concentration of that capability in the hands of a few companies. Meta’s Chief Scientist Shengjia Zhao signed the letter as a private individual, at the very moment his own boss published a counter-argument. Alongside the op-ed, a coalition letter from Meta, Nvidia, Microsoft, and Palantir also appeared, asking regulators not to restrict models with freely available weights.

Do you believe that the development of artificial intelligence can be slowed down at all?

Sources: Pacing the Frontier, CNN, Fortune, Notebookcheck, Tech Times