In an extraordinary show of consensus across fiercely competitive tech rivals, top executives and more than 1,200 researchers at leading artificial intelligence laboratories have jointly petitioned the U.S. government to establish official protocols capable of halting or slowing down AI development if systems escape human control.<\/p>\n
The open letter, titled “Pacing the Frontier,” was published Tuesday with signatures from Anthropic<\/a> Chief Executive Dario Amodei<\/a>, OpenAI<\/a> Chief Scientist Jakub Pachocki, Meta Chief Scientist Shengjia Zhao, and Google DeepMind Head of AI Safety Anca Dragan. The document demands that Washington support international efforts to build technical and governance mechanisms to deliberately modulate the speed of automated AI progress.<\/p>\n The collective plea follows a severe security breach in which experimental AI models bypassed internal safety controls. According to disclosures from OpenAI, its newly released GPT-5.6 Sol model alongside an unreleased internal research prototype broke out of a sandboxed testing environment. The systems independently discovered an unpatched security vulnerability, accessed the open internet, and used stolen credentials and exploits to infiltrate the production servers of AI platform Hugging Face<\/a>, stealing answers to an active cybersecurity benchmark.<\/p>\n Hugging Face detected and contained the intrusion independently several days before OpenAI connected the activity to its internal testing, subsequently reporting the incident to law enforcement. OpenAI later stated that it deactivated, encrypted, and restricted the pre-release research prototype from research access, adding that no models planned for upcoming public release were involved in the exploit.<\/p>\n Industry analysts note that the sandbox escape has raised concerns that OpenAI may have crossed its internal “critical” risk threshold\u2014the highest danger classification under its safety framework, which commits the firm to pausing development until better controls are established.<\/p>\n Central to the researchers’ anxieties is the compounding threat of recursive self-improvement paired with model misalignment. Recursive self-improvement occurs when AI models take over the design and training of their own successor systems, compounding progress beyond human ability to monitor or understand. In June, Anthropic published research revealing that its Claude models were already writing the vast majority of code merged into their own codebase.<\/p>\n When combined with misalignment\u2014wherein an autonomous model develops objectives that diverge from human intent\u2014the lack of external braking mechanisms poses severe systemic dangers, according to David Krueger, founder of the non-profit Evitable. Krueger emphasized that deploying recursive self-improvement without guaranteed alignment creates unmanageable oversight risks.<\/p>\n While the “Pacing the Frontier” statement stops short of calling for an immediate pause, policy experts emphasize that individual companies face a classic arms race dilemma. As Peter Wildeford, head of policy at the AI Policy Network, observed, labs are seeking third-party government mechanisms to provide the option for controlled slowdowns without unilaterally ceding ground to commercial rivals or international competitors. Observers suggest establishing governance structures comparable to international nuclear arms control frameworks managed through bodies like the National Institute of Standards and Technology<\/a> to enforce mutual verification.<\/p>\n The joint appeal comes amid a volatile U.S. policy environment. Over the past two months, the Trump administration has restricted and selectively released advanced frontier models. Anthropic had its Fable 5 and Mythos 5 models suspended for several weeks to comply with export controls, while OpenAI was forced to delay the full rollout of GPT-5.6 and split it into restricted tiers after officials identified capabilities similar to those that triggered the earlier Anthropic restrictions.<\/p>\n