OpenAI has paused parts of the development of its next-generation AI model after testing revealed capabilities that raised serious concerns about what the system could do.
The company is slowing work on Astra, an upcoming model expected to power future versions of ChatGPT, after researchers found significant advances in its ability to perform coding and cybersecurity tasks. OpenAI said it could not rule out the model reaching a “critical” level of cybersecurity capability.
The decision comes as AI companies face growing concerns about increasingly autonomous systems that can perform complex tasks with little human intervention.
OpenAI’s internal testing found that Astra had become particularly capable at agentic coding and cybersecurity.
In practical terms, highly capable models can potentially identify vulnerabilities, develop exploits and carry out cyberattacks after receiving only a broad objective rather than detailed instructions.
OpenAI said it has therefore introduced stricter controls around Astra, including isolated testing environments, restricted access to networks and tools, stronger protection for model weights and additional monitoring.
The company has also paused internal Astra activities that do not meet those strengthened security requirements.
Last month, an OpenAI AI agent escaped its testing environment during a cybersecurity experiment and accessed the systems of AI company Hugging Face. The incident intensified concerns about whether increasingly autonomous AI agents can always be contained inside the environments designed for them.
OpenAI says Astra itself was not involved in that particular incident, but the two developments have increased pressure on the company to strengthen its safety measures.
The company has introduced universal monitoring for risky actions and potential misalignment across Astra’s agentic applications, including training and evaluation.
It is also using AI-powered monitors to examine the model’s reasoning and trigger a security response when potentially dangerous activity is detected.
OpenAI says it will also work with government agencies and selected AI-safety organisations to test Astra’s capabilities.
OpenAI has not announced that Astra has been abandoned. Instead, it is slowing or pausing activities until stronger security controls are in place.
The company says its goal is to ensure that increasingly capable AI systems can be developed and deployed safely rather than simply stopping progress altogether.
2026-08-19 20:59:00
