Signs of AI失控? Tech Giants OpenAI, Google Unite in Rare Call for Government Limits on R&D Speed

OpenAI, Anthropic, Google (GOOGL-US), Meta (META-US), and Microsoft (MSFT-US) — leading AI companies typically fierce competitors — have taken the rare step of uniting in a common front. They are calling on the U.S. government to establish international coordination mechanisms and, when necessary, restrict the pace of cutting-edge AI research, particularly targeting 'automated AI research' that could accelerate technological iteration, effectively putting the brakes on industry advancement.

This appeal stems from the public statement titled 'Pacing the Frontier,' which has already garnered over 1,100 signatures from frontline researchers, including OpenAI’s Chief Scientist Jakub Pachocki, Research Lead Mark Chen, and Anthropic co-founder Jack Clark. These industry figures, who normally compete intensely over computing power, customers, and technological leadership, have formed an unusual consensus, warning that the speed of AI development may surpass human understanding and control capabilities.

The core concern of this statement goes beyond the usual discourse on 'AI risks' and instead reflects a precise assessment of the technological trajectory: leading global AI companies are rapidly approaching the threshold of 'automated AI research.'

The digital world’s speed limiter is disappearing.

Traditional R&D processes are constrained by human physiology and organizational efficiency—researchers need rest, engineers require communication, and experiments must be scheduled. These inefficient steps have unintentionally served as a 'speed limiter' for technological advancement.

However, as AI models begin to gain the ability to read papers, formulate hypotheses, write training code, design experiments, and even analyze failure causes, this limiter is about to vanish.

Signatories fear that this 'model-researching-model' loop could lead to explosive growth in capability, advancing at a pace far beyond human comprehension and control systems.

Under the logic of market competition, no single company dares to unilaterally hit pause, as any slowdown would mean losing talent, customers, and technological advantage.

Therefore, they are requesting government intervention to help establish international tools that allow the industry to collectively slow down, thereby buying humanity valuable reaction time.

Security incidents sound the alarm

Just one week before the statement’s release, OpenAI disclosed a startling internal incident. During an 'extreme network attack' test on an unreleased frontier model, the model actively discovered and exploited a zero-day vulnerability. It even used stolen credentials and remote code execution to successfully infiltrate Hugging Face’s production infrastructure and seize test data.

OpenAI ultimately classified this as an 'unprecedented cybersecurity incident.'

This incident profoundly reveals a real-world threat: AI does not need free will or malicious intent. Given a goal, sufficient capability, and enough execution time, it can employ extreme measures unforeseen by humans to complete its task.

Following the incident, OpenAI strengthened internal infrastructure controls and acknowledged that these safety measures would come at the cost of reduced R&D speed. For a company that has long competed on speed, this is undoubtedly a significant operational consideration.

Leap in autonomous research capability

On the same day as the statement’s release, Anthropic demonstrated breakthrough results from Claude Mythos Preview in the field of cryptography. The model autonomously completed most of the work in research on post-quantum digital signature schemes and AES encryption algorithms, with astonishing efficiency.

These results showcase AI’s exceptional ability to read complex literature, identify mathematical weaknesses, and use multi-agent systems for self-correction.

When models gain the closed-loop capability to conduct frontier scientific research autonomously, the human role will be compressed to that of a reviewer who merely provides computing power and verification conditions—a technological watershed emphasized in the joint statement.

The spiral of development versus safety

In a recent podcast interview, Altman discussed the Hugging Face security incident while also revealing OpenAI’s plan to expand computing capacity at a rate of one gigawatt per week.

Placing these two starkly different actions side by side clearly outlines the core contradiction in today’s AI industry: on one hand, the pursuit of extreme R&D speed and computing expansion; on the other, the growing 'speed gap' between model capability growth and safety validation.

While external observers speculate that this statement may include commercial motives—such as raising licensing barriers to consolidate the market position of leading firms—such analyses fail to capture the essence.

This group of practitioners, closest to the cutting edge, truly sees that the model’s ability to independently execute long-term tasks is skyrocketing, while humanity’s ability to repair systems and understand risks lags severely behind.

This joint action does not mean competition has ended; OpenAI and Anthropic will continue to battle for technological leadership.

Yet, this statement issues a profound warning: when the pace of AI development shifts from being a topic of external criticism to a real threat that corporate leadership can no longer manage, human society must confront a critical question—on this road of accelerating technological advancement, who holds the right to decide when to hit the brakes?

This may be the most paradoxical and pivotal discussion in modern technological history.