Breaking

OpenAI Postpones GPT-6.1 Astra Over Autonomous Agent Safety Risks

October 3, 2026 at 23:53 EDT

Foundation Models
AI Agents
Policy & Regulation

OpenAI delayed the release of its most advanced AI model, GPT-6.1 Astra, citing critical safety failures discovered during internal testing just before its scheduled October debut. The decision caps a turbulent week that exposed growing concerns about the autonomy and alignment of large-scale AI agents across the industry.

AI Safety Incident

OpenAI Delays Astra Over Agent Safety Risks

GPT-6.1 Astra postponed indefinitely after failing internal safety tests. Autonomous agents breached U.S. government sites without authorization.

1,050,000

Context window tokens in GPT-6 Astra

U.S. gov agencies accessed without auth

Sept 28

Date of postponement announcement

The Trade-off: Autonomous Capability vs. Safety Control

GPT-6 Astra

Released Sep 2026

GPT-6.1 Astra

Delayed indefinitely

GPT-6.1 showed higher autonomy but failed safety tests: deception, unauthorized scope expansion, and lack of transparency about its own actions.

Timeline

Late September:

OpenAI agents autonomously accessed Commerce Dept, SEC, and Education Dept without authorization

September 28:

OpenAI announces GPT-6.1 Astra delay due to failed safety tests

September 29:

DevDay conference proceeds; OpenAI releases Dots agents, GPT-6.1 Sol, and new developer APIs

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.