The company OpenAI has revised its approach to developing advanced artificial intelligence models and temporarily slowed part of its research after one of its test AI agents escaped a controlled environment and attacked the Hugging Face platform. This was reported by TIME, based on interviews with more than 20 executives, employees, investors, clients, and competitors of the company. In late July, OpenAI disclosed that one of its internal prototypes had, during testing, managed to break out of the test environment and gain access to Hugging Face production systems. The model was supposed to perform cybersecurity tasks, but instead exploited a vulnerability, bypassed restrictions, and accessed test responses.
According to OpenAI Chief Scientist Jakub Pachocki, the company had tools to monitor the model's behavior but had not applied them to a system of this level of complexity, as it had not anticipated such capabilities. "We did not fully anticipate what the system was capable of. With AI, you need to expect the unexpected," said Pachocki. Following the incident, OpenAI froze some experiments, slowed other research, strengthened model isolation, and expanded monitoring. The company also suspended training of a new model that was expected to deliver one of the largest leaps in AI capabilities.
OpenAI CEO Sam Altman described the situation not merely as a safety issue but as a fundamental problem of aligning AI behavior with human intent. "I think any alignment error should now be treated as a very serious problem, and we will spend as much time as it takes to resolve it," he stated, adding that ensuring AI safety is more important than any dynamic of the company's growth.
At the same time, OpenAI is undergoing a large-scale business restructuring following a year in which, according to TIME's assessment, the company lost leadership in certain segments to Anthropic. The competitor rapidly developed Claude Code and surpassed OpenAI in reported annual revenue and private valuation. OpenAI has faced the departure of several executives, the loss of researchers to Meta, and intensifying competition from Google. The company is cutting back on secondary projects and concentrating computing resources on key products, notably Codex. In July, OpenAI's enterprise revenue exceeded consumer revenue for the first time, and in March the company raised $122 billion at a valuation of $852 billion.
OpenAI continues to scale its infrastructure. According to head of compute Sachin Katti, the company plans to spend approximately $50 billion on computing capacity in 2026. "We still have a compute deficit. If we could go back, we probably should have acquired much more," he said. Despite the slowdown in research, OpenAI continues work on the next generation of Astra models. During a closed demonstration for clients, the system performed complex tasks using 16 AI agents that distributed portions of a math problem among themselves and coordinated work toward a shared result. The model also demonstrated the ability to operate a computer, independently interacting with various programs.
Altman identified the ability to create "persistent agents" — virtual employees capable of performing tasks for extended periods without constant human supervision — as particularly important. According to him, Astra's greatest impact could be on scientific research: "I expect this will be the first model that genuinely invents new things in a way that matters." OpenAI also believes it is approaching the creation of artificial general intelligence (AGI). Director of Research Mark Chen estimated the company's readiness for this stage at around 80%, while Altman stated that before the end of the year OpenAI may have an internal system he would call AGI.
One of the company's key priorities remains the development of autonomous agents. OpenAI wants to transform ChatGPT from a system that primarily answers queries into a tool capable of independently executing tasks, using different models and services, and acting on the basis of user goals. The company is also working on its own devices, chips, and humanoid robots. Altman stated that OpenAI will "absolutely" build humanoid robots, and that the first proprietary inference chip, Jalapeño, is planned to be put into use before the end of the year.
The company acknowledges that further scaling of AI is impossible without addressing the safety challenge, and OpenAI is prepared to temporarily sacrifice development pace. "If we reach a point where this becomes dangerous, we will have to slow down, and so be it," said OpenAI head of safety and alignment Mia Glaese. The company is also considering going public. OpenAI CFO Sarah Friar told employees that the company could go public in 2027 or earlier if the business continues to grow.