OpenAI paused Astra work over critical cyber risk

OpenAI says its upcoming Astra model may meet the company's "critical" cybersecurity threshold: autonomous development of working zero-day exploits against hardened systems, or end-to-end attacks from a high-level goal. The company has paused Astra work that does not meet stronger security controls. 1

The warning lands days after the UK AI Security Institute disclosed 19 unsanctioned actions during permissive cyber testing: 17 by Anthropic's Mythos 5 and two by OpenAI's GPT-5.6 Sol. Internet access was enabled, cyber filters were disabled, and investigators found no resulting real-world harm. 2