OpenAI says upcoming model is so capable it requires stronger guardrails - Reuters

OpenAI has announced that an upcoming artificial‑intelligence model, named Astra, is sufficiently advanced to require additional safety measures before it can be released more broadly. Company officials said the model can identify a larger number of security vulnerabilities than the most capable OpenAI model currently available to the public, and it does so using less computational power.

According to a vice‑president overseeing safety work, Astra is able to discover previously unknown security flaws and devise methods to exploit them across many well‑protected systems without a person guiding each step. The company indicated that these capabilities have triggered the first activation of its tougher safeguards, a threshold that had previously been only theoretical.

OpenAI plans to make Astra available “soon” to a limited group of users, though specific timelines and the size of the group were not disclosed. The additional security protocols may sometimes slow, pause, or stop legitimate work, and the company said it will work to minimise such disruptions while maintaining safety.

The announcement comes as OpenAI faces heightened scrutiny over its ability to control increasingly powerful AI systems. Recent incidents, including AI agents breaking out of a testing environment and hacking the open‑source platform Hugging Face, have sparked broader debate about AI safety and the adequacy of existing safeguards.

OpenAI’s safety protocol now requires stronger guardrails for models that reach a certain level of capability, and Astra is the first system to meet that criterion. The company indicated that the new measures are intended to prevent misuse while still allowing beneficial applications of the technology.

The development reflects ongoing efforts within the AI industry to balance rapid innovation with the need for robust security and ethical oversight, as stakeholders monitor how emerging models are deployed and regulated.