OpenAI Hits Pause on New AI Model After It Shows Alarming Cybersecurity Skills

OpenAI has announced it is pausing certain internal activities related to its upcoming AI model, known as Astra, after internal testing revealed the model had made significant advancements in agentic coding and cybersecurity capabilities. These skills, while impressive from a technology standpoint, raised concerns serious enough that OpenAI decided to slow down and reassess before continuing further development or deployment.

In response to these findings, OpenAI says it is now implementing stronger security controls for its higher-capability models, including isolating certain systems and activities to reduce risk. The move highlights a broader industry concern: as AI models become more capable at writing and analysing code, they could potentially be used to discover vulnerabilities, write exploits, or automate cyberattacks far more efficiently than before — a capability that could be misused if it fell into the wrong hands or was deployed without adequate safeguards.

While Astra is still under development and not yet publicly available, this pause signals that even leading AI companies are recognising the dual-use nature of advanced AI: the same skills that make a model a powerful coding assistant could also make it a powerful hacking tool. For small businesses, this is a reminder that the cyber threat landscape is evolving quickly as AI tools grow more sophisticated on both the defensive and offensive sides.