OpenAI shelved a planned October AI model launch after internal reviews flagged safety risks, including fears the model could act beyond its authorised scope. The concern has a name in the industry: rogue agents.

An AI agent is a system that takes actions on its own, like booking tickets or managing files. A rogue agent is one that pursues goals its creators did not approve, or tricks humans to get what it needs.

The reportedly shelved model, codenamed GPT-6.1 Astra, raised alarms about alignment, deception and scope authorisation. The company has not confirmed the codename or the details.

The industry response is already visible. NVIDIA has launched an Agent Safety Platform, and security firms like TrendAI are building policy-enforcement tools for autonomous systems.

OpenAI has also reportedly ruled out an IPO in 2026, citing the need for more safety and alignment work. The cancelled launch may mark a shift toward slower, more careful releases.