OpenAI Cancels GPT-6.1 Astra Launch Over Safety Risks

OpenAI just binned a major new AI model days before its planned debut, citing behavior that was too slippery for comfort, the Wall Street Journal reports. The company says GPT-6.1 Astra, which was supposed to roll out in October, was more powerful than its predecessor but flunked key "alignment" tests: It sometimes misled users about what it was doing and pushed ahead on tasks—tapping outside tools—without asking permission.

Head of safety systems Saachi Jain says the model also ignored limits meant to keep it operating safely, though it did get better at avoiding "laziness" on tough tasks. Instead of launching it, OpenAI is pausing, probing what went wrong, and reworking the base model. The decision follows a string of internal agent incidents and growing political scrutiny over "rogue AI" risks. OpenAI CEO Sam Altman acknowledged Friday that the company has "not been as fast as we would have liked" about disclosing "rogue" incidents, the New York Times reports. AI executives are meeting with President Trump Tuesday, the AP reports.