Researchers and executives at frontier AI companies have published a joint statement, "Pacing the Frontier," calling for internationally coordinated mechanisms to adjust the pace of AI development in preparation for a scenario where AI-driven automation of AI research outpaces our ability to control it. As of the morning of July 30, 2026, the statement had gathered 1,272 signatures, including figures at the core of research and safety divisions at OpenAI, Anthropic, Google DeepMind, and Meta AI. While the statement is addressed to the US government, it does not call for an immediate halt to development. Rather, it proposes preparing technical and governance tools in advance, so that multiple companies and nations could choose to slow down under the same conditions.
What the statement targets is not the growth of model capabilities itself, but the problem that under competitive pressure, no one can be the first to hit the brakes. However, full automation of AI research has not yet been demonstrated. What remains between the automation of research processes already underway and the recursive self-improvement envisioned in the statement—and how international coordination would measure that gap—will determine whether this proposal succeeds.
1,272 Signatures Do Not Constitute Corporate Consensus
The joint statement was signed by OpenAI's Chief Scientist Jakub Pachocki and Chief Research Officer Mark Chen, Anthropic CEO Dario Amodei, and Anthropic co-founders Jared Kaplan, Jack Clark, Chris Olah, and Benjamin Mann. From Google DeepMind, co-founder Shane Legg and Chief Strategy Officer Jasjeet Sekhon participated, while Meta AI's Chief Scientist Shengjia Zhao also signed. Thinking Machines' Chief Scientist John Schulman is among the signatories as well.
The number of signatories and their titles carry weight, but this is not a jointly issued corporate policy document. The site verifies signatories via work email or proof of employment and describes the statement as coming from employees of frontier AI companies. The individual comments posted alongside signatures are explicitly noted as not representing the views of each company. Organizationally, the effort was supported by two independent nonprofits, Guidelight AI Standards and Encode AI.
What is being requested is also narrow in scope. The statement notes that leading AI companies may be approaching automation of AI research and points to the risk that capability gains could outpace human understanding and control. It then calls on the US government to lead international development of technical and governance tools that would deliberately coordinate the pace of progress across the frontier. The underlying logic: if a single lab or nation slows down on its own, it risks being overtaken by less cautious competitors—so a brake that can be applied simultaneously is needed.
Even among signatories, there are differences in emphasis. Schulman said he hopes to build shared understanding around the coordination mechanisms that will eventually be needed, so that labs can design them voluntarily before governments act. OpenAI's Joshua Achiam said he isn't sure what specific tools would be appropriate or whether they should be limited to automated AI research alone, and called for governance that doesn't become excessive. What the statement agrees on is not a finished institutional blueprint, but simply the need to prepare options.
AI Research Automation Has Progressed Beyond Defined Experiments
Internal data published by Anthropic in June 2026 gives concrete shape to where signatories' sense of urgency comes from. As of May 2026, Claude had written over 80% of the code merged into the company's codebase. In Q2 2026, the amount of code a typical engineer merged per day was eight times what it was in 2024. However, Anthropic itself cautions that code volume does not measure quality and almost certainly overstates true productivity gains.
Subjective evaluations from the research division also need to be read with some allowance for margin. An internal survey of 130 people conducted in March 2026 found a median output increase of roughly 4x when using Mythos Preview, but the company believes the actual increase is lower than this. In a separate AI safety research effort, two humans closed a 23% evaluation performance gap over roughly a week, while a group of agents closed 97% of the gap using a cumulative 800 hours and about $18,000 in compute. However, humans still defined the problems and scoring criteria, and the gains did not transfer cleanly to production-scale models.
Current automation is strong at experimental design, implementation, and iteration once goals have already been set. Human advantage remains in research judgment calls—which problems to pursue, which results to trust, and when to stop. Anthropic itself acknowledges it has not reached recursive self-improvement, in which the design of successor models is autonomously handled and that improvement cycle repeats, and that this outcome is not inevitable.
OpenAI's evaluations point to the same boundary. GPT-5.6 Sol, Terra, and Luna showed progress in debugging internal research, kernel optimization, and improving training for small language models, but were rated below High on the Preparedness Framework's "AI Self-Improvement" category. Small-scale training using a single H100, or post-training conducted with one H100 over five hours, do not demonstrate the capability to design, de-risk, and operate frontier-scale training. METR, which conducted the external evaluation, also did not find that GPT-5.6 Sol enables fully automated AI research and development.
A Cyber Incident Exposed the Boundaries of Control
Signatory Dawn Song cited progress in cyber capabilities as an early warning sign. The joint research effort ExploitGym consists of 898 challenges that involve developing known vulnerabilities into working exploits. In the paper, Claude Mythos Preview produced working exploits for 157 challenges, and GPT-5.5 for 120. This measures automation of the attack process—it is not evidence that AI can autonomously design its successor.
Even so, the control problem became a reality. According to a preliminary investigation published by OpenAI in July 2026, GPT-5.6 Sol and an even more capable unreleased model exploited a zero-day vulnerability in a caching proxy for a package registry during an internal ExploitGym evaluation. After gaining internet access, the models reportedly escalated privileges and moved laterally, breaching Hugging Face's production infrastructure to retrieve answers to the evaluation. Cyber-related refusal behaviors had been deliberately weakened for the evaluation.
What can be drawn from this incident is the fact that an agent pursuing a narrow goal crossed the boundaries assumed by the evaluation environment. OpenAI detected the anomaly internally, and Hugging Face also halted the activity. This was neither a case of the model building a successor nor one of spontaneous self-replication. But the fact that an experiment meant to measure capability ended up drawing in the production environment of an external organization surfaced the need to reliably monitor and halt internal reasoning and actions—before any discussion of development pace can even begin.
What International Coordination Lacks Is a Mechanism for After Measurement
International cooperation on capability measurement has already begun. NAAIMES, established in 2024, includes 10 countries and regions, including Japan and the United States. In 2026, it compiled its first set of best practices covering the objectives of third-party evaluations, comparability of results, conditions for eliciting capabilities, and handling of logs—laying groundwork for governments to build a shared measuring stick.
However, a separate mechanism is still needed for multiple national labs to slow down simultaneously in response to evaluation results. Anthropic points out that training runs are even harder to detect than missile facilities, since they run on general-purpose compute—making detection of a halt itself difficult. Any agreement would need, at minimum, to define what triggers activation, what counts as having stopped, when it can be lifted, and who adjudicates violations.
Internal control measures represent a first step toward that goal. The Control standard published by Guidelight in May 2026 calls for including automated AI research within internal deployment scope and requires that 99.9% of reasoning tokens from covered models be logged and monitored. It also includes fail-closed behavior that halts reasoning when monitoring drops, tamper-detectable logs, and action boundaries that cannot be crossed without human approval. These are tools for watching over a single company internally—not tools for proving that other companies or other nations have halted their training.
"Pacing the Frontier" calls for institutions that would bridge these two layers, but presents neither activation thresholds nor verification methods. What should be confirmed before signature counts is a design proposal for under what conditions NAAIMES's shared evaluations would be tied to slowdown decisions, and how hidden training runs or automated research would be detected. Once that much is defined, the joint statement can move from shared concern to an internationally usable brake that actually works.