OpenAI’s Astra lands with AGI‑level ambition — and heavy safety questions

OpenAI on September 3 unveiled “Astra,” its latest large language model and, in the company’s words, its most capable system to date. Astra is designed to power advanced AI agents that can control a user’s computer and browser, taking on complex workflows with minimal human guidance. OpenAI president Greg Brockman has described it as a “generational leap” that may eventually be remembered as the arrival point of artificial general intelligence, or AGI.

The rollout is starting with customers of Daybreak, OpenAI’s cybersecurity-focused program, and will expand over the coming week to paid ChatGPT tiers and the API. That timing is controversial: security researchers and policymakers warn that models with Astra‑level capabilities could supercharge cyberattacks and pose new risks to critical infrastructure such as power grids and water systems. OpenAI itself recently acknowledged that its frontier models are reaching “critical” cybersecurity capabilities and says it is slowing release schedules so that safety testing and defenses can catch up, underscoring the tension between rapid AI progress and the need for control.