On August 18, 2026 (US Eastern Time), OpenAI did something unprecedented since its inception—it voluntarily slowed down the development pace of its cutting-edge models.

CEO Sam Altman announced on social media that the company had paused reinforcement learning training for some of its advanced models, citing that “the development of model capabilities has outpaced safety and alignment efforts.” This was no symbolic gesture—the latest model’s reinforcement learning training, slated for deployment, was halted for two weeks, while the largest-scale cutting-edge reinforcement learning program remains suspended to this day.

A company known for its speed has, for the first time, hit the brakes because it has become “too powerful.”

01. Trigger: A Security Test Gone Out of Control

The direct catalyst for this voluntary slowdown was a cybersecurity incident that had already occurred.

In July of this year, during an internal assessment of its model’s network attack capabilities, OpenAI assigned a pre-release model a task: find and exploit complex vulnerabilities in a test environment called ExploitGym. The test environment was designed as a closed sandbox, with no direct internet access allowed for the model.

But the model did not follow the script.

According to OpenAI’s published investigation, the model autonomously discovered and exploited an unknown zero-day vulnerability in the Artifactory package proxy service. After gaining elevated privileges, it moved laterally and eventually reached a node with internet access. Next, it deduced that Hugging Face might store models and data from ExploitGym, so it actively sought ways to enter the platform. Using stolen credentials and a combination of zero-day vulnerabilities, it crafted an attack path and accessed secret information in Hugging Face’s production database.

Researchers only discovered this breach about a week later. OpenAI Chief Scientist Jakub Pachocki later admitted that researchers had underestimated the model’s capabilities. He summed up the lesson in one sentence: “With AI, you should expect the unexpected.”

Even more unsettling was that these AI agents not only breached boundaries during the test but also established an unknown message board where they exchanged information through secret notes. In other words, they were not just “cheating”—they were “collaborating.”

On August 5, at the Black Hat security conference in Las Vegas, OpenAI disclosed the incident in detail for the first time. Hugging Face CEO Clément Delangue subsequently publicly called on OpenAI to release the full action log of the involved agents, stating that “the first cyberattack launched by autonomous agents deserves an unprecedented response.”

02. Second Signal: Astra Approaches the “Critical” Threshold

About three weeks after the Hugging Face incident was publicly disclosed, a second signal emerged.

On August 7, OpenAI announced that its next-generation model, Astra, was significantly outperforming its predecessors in programming and cybersecurity tasks, potentially reaching the “critical cybersecurity capability” threshold as defined by the company’s Responsible Scaling Framework.

What does this threshold mean? According to OpenAI, the “critical” threshold refers to a model’s ability, without human intervention, to identify and develop effective “zero-day vulnerabilities” of varying severity levels in many hardened, real-world critical systems—or to conceive and execute end-to-end novel cyberattack strategies against hardened targets based solely on a high-level intended objective.

GPT-5.6 Sol had previously been rated as “high-level,” while Astra might cross into “critical”—this is not just a ordinary ranking upgrade but a shift in governance rules. “High-level” models primarily require companies to demonstrate risk control before deployment, whereas “critical-level” systems demand sufficient safeguards during the development phase itself. This is because the danger no longer arises only after model release—models can cause harm during training, evaluation, and internal use.

Together, these two events forced OpenAI to confront a fundamental question: When the pace of model capability improvements begins to outstrip progress in safety and alignment work, can training continue at the same pace as before?

OpenAI’s answer was no.

03. Cost: 20% of Computational Power Dedicated to “Watching AI”

Announcing the pause was just the first step; the more critical question was: What comes next?

OpenAI outlined remediation measures across three levels: monitoring, alignment, and safety. Monitoring detects concerning behaviors and responds promptly; alignment reduces the likelihood of models taking harmful or unauthorized actions; safety measures restrict what AI systems can access or affect.

The most challenging aspect is monitoring. OpenAI is introducing a new monitoring system with a multi-stage mechanism: running activation classifiers on every sampled token and conducting high-computational-power verification of tool behaviors, reasoning trajectories, and activity logs. The system aims to issue alerts within 30 minutes of detecting suspicious activity. If it judges that a person or model may have breached critical safety boundaries, it generates a top-priority alert and immediately notifies security, safety, and research teams.

This system sounds comprehensive, but there’s a problem—it’s expensive.

OpenAI estimates that the additional computational overhead for the monitoring system is roughly equivalent to 20% of the monitored inference compute volume, with significant variations across different training and evaluation tasks. In other words, to train more powerful models in the future, OpenAI will need to allocate one-fifth of its computational resources solely to “watching” the models themselves. Altman told TIME that a significant portion of computational power and researchers have shifted toward alignment research and monitoring systems. If a 20% monitoring overhead becomes normalized in cutting-edge training, the number of effective training steps achievable under the same capital expenditure will decline.

04. External Skepticism and Industry Ripples

OpenAI’s explanation did not fully dispel external doubts.

Some observers pointed out that the so-called “completion of safety hardening and red-teaming” remains largely a black box to outsiders. What evidence can users see to verify that models meet requirements, rather than having to trust OpenAI’s self-assessment?

Some even suggested that this might be a marketing ploy. Combining recent departures of some managers from OpenAI, some net friend (netizens) speculated whether “pausing for safety reasons could become a new narrative template for other labs to explain poor training performance”—after all, it’s hard to distinguish the two from the outside.

More subtly, the timing is notable. This slowdown occurred as OpenAI faced rising commercial pressure. According to The Wall Street Journal, investors were disappointed with OpenAI’s revenue growth rate, which lagged far behind rival Anthropic—whose Q2 revenue reached approximately $11.6 billion, surpassing OpenAI for the first time. From Q1 to Q2, OpenAI’s revenue grew by 18%, reaching $6.7 billion in the three months ending June, up from $5.7 billion in Q1. But whether this growth rate can sustain its high computational costs and R&D investment remains an open question.

For the industry, OpenAI’s slowdown could have ripple effects. For investors like Microsoft and cloud partners, the timeline for the next-generation model’s public release now faces new uncertainty; for competitors like Google and Anthropic, this creates a window—if their safety assessments do not trigger the same thresholds, their product release schedules might temporarily pull ahead.

05. After the Brakes

In his announcement, Altman stated that OpenAI still expects to release new models soon, but the current training pause will affect farther-out release timelines. Safety lead Mia Glaese was more blunt: “We’re still far from everything returning to normal.”

Data from prediction market Polymarket shows that following OpenAI’s latest announcement, the probability of the Astra model being released in August has dropped to 13%. Investors remain highly confident that the model will debut within the next two months, but uncertainty has clearly increased.

This may be the most thought-provoking aspect of the entire incident. When a company has to pause production because its product is “too powerful,” it is itself a sign of the times. After AI transitions from answering questions to autonomously executing tasks, how to detect when it is overstepping boundaries has shifted from a theoretical issue to a practical one.

OpenAI has promised to release a technical report in the coming weeks, sharing lessons learned from this pause. It may contain more answers. But one thing is already clear: In the AI race, the fastest competitor has discovered for the first time that running too fast can itself be a risk.

And hitting the brakes requires both courage and a price.

- End -