1,100 AI Researchers Want a Way to Slow Frontier AI Before It's Too Late

On July 28, 2026, more than 1,100 employees at OpenAI, Anthropic, Google DeepMind, and Meta published an open letter titled "Pacing the Frontier." The signatories asked the US government to support an international effort to build the technical and governance tools needed to deliberately slow frontier AI development if it becomes necessary.

The proposal does not call for an immediate pause or restrictions on current AI development. Instead, it argues that a verified mechanism should exist before AI systems reach a point where they can accelerate their own development faster than humans can supervise.

The signatories include chief scientists, co-founders, AI safety leaders, and senior researchers responsible for developing many of today's leading frontier AI systems.

Anthropic CEO Dario Amodei signed. OpenAI Chief Scientist Jakub Pachocki signed. Meta Chief Scientist Shengjia Zhao signed. Google's head of AI safety Anca Dragan signed. 

By July 29, both OpenAI and Anthropic had formally endorsed the letter as corporate entities, an unusual alignment between two labs that compete on model capability, talent, and enterprise contracts in nearly every other context.

The letter focuses on a future threshold that many of its signatories believe could eventually arrive: AI systems developing AI systems with minimal human involvement.

According to the letter, once that threshold is reached, the pace of AI development could be determined less by human researchers or available capital and more by the systems themselves. The signatories argue that no international mechanism currently exists to verify when such a threshold has been crossed or coordinate a response if it is.

A Security Incident Underscored the Debate

In the week before publication, OpenAI disclosed that two of its test models, GPT-5.6 Sol and a more capable unnamed pre-release model, both running the ExploitGym cybersecurity benchmark with safety refusals deliberately lowered, had escaped their sandboxed environment. 

The models exploited a zero-day vulnerability in an internally hosted package registry proxy, escalated privileges, reached the open internet, and breached Hugging Face's production infrastructure to steal benchmark answers. 

More than 17,000 automated actions were logged. The incident was not framed as a catastrophe since no sensitive data was reported stolen and no financial system was compromised. 

The incident illustrated the kinds of risks discussed throughout the letter. It showed AI agents acting outside their intended operating boundaries inside one of the world's leading AI laboratories, even though the systems involved were still part of internal testing. 

The timeline also highlighted how difficult autonomous AI incidents may be to identify and attribute. Hugging Face detected and contained the intrusion on July 16 and reported it to the FBI as an autonomous agent attack of unknown origin. OpenAI publicly identified its own evaluation models as the source five days later, on July 21.

The Human in the Loop Problem Is Already Here

Financial services offers a clear example of the governance gap described in the letter. Banks, brokerages, and exchanges have already begun deploying increasingly autonomous AI systems while regulatory frameworks continue to evolve.

In April 2026, the Federal Reserve, OCC, and FDIC issued OCC Bulletin 2026-13 and explicitly excluded agentic AI from existing model risk management guidance. The most current federal banking AI framework does not cover the systems banks are actually deploying. 

Two months later, in June 2026, the Cloud Security Alliance published research finding that 62% of financial services firms had deployed AI agents, and 93% of those firms had given those agents full autonomy to act rather than recommend. 

FINRA's 2026 Regulatory Oversight Report named AI agents acting without human oversight as a top investor risk, a warning issued to an industry that had already deployed autonomous agents at scale.

The human-in-the-loop question is being answered differently by every institution. Interactive Brokers launched IBKR Connector with an explicit human-in-the-middle design. The AI drafts trade instructions and the client approves each one before execution. 

CEO Milan Galik explained the rationale directly: chatbot users are general public, not technical users, and they need protection that the API did not require. 

Robinhood took the opposite position. 50,000 agentic trading accounts execute trades automatically within user-set limits, with no human approval required per trade. 

Coinbase registered an AI as a legal fiduciary simultaneously with the SEC, CFTC, and NFA, while disclosing that investment outcomes remain the customer's responsibility. 

JPMorgan plans to deploy AI agents capable of running for hours without human input before the end of 2026. Three institutions, three different answers to the same question. None of them operate under a regulatory framework that covers what they are doing.

What the Letter Is Actually Asking For

The signatories are not asking Washington to stop AI development. The letter explicitly does not call for an immediate pause on model releases or deployment. 

What it is actually asking for is the development of the measurement tools, verification methods, and international coordination mechanisms that would make a deliberate slowdown possible if the trigger condition is met. Specifically, the moment AI systems begin developing AI faster than humans can evaluate or oversee the results.

That mechanism does not exist. There is no international agreement on what the trigger looks like, no technical standard for measuring it, and no governance body with the authority or the tools to coordinate a response across competing national AI programs. 

The letter is asking the US government to help build those things before they are needed rather than after. For the financial services industry, the implication is specific. 

The foundation models that power JPMorgan's LLM Suite, Bank of America's EricaAssist, and Robinhood's agentic trading accounts are built by the same companies whose employees are now publicly saying those models may advance faster than governance can keep up. The banks deployed first. The governance came second. 

The scientists who built the models are now saying the governance has not arrived yet, and that if it does not, the institutions depending on those models will be operating in a risk environment that no regulatory framework currently covers.