OpenAI is rushing to fix security issues in its next major model before launch, according to The Information. This isn’t the first time. At the end of September, OpenAI itself delayed the release of GPT-6.1 Astra—originally scheduled for October—because it performed worse in security tests than GPT-6 Astra. According to OpenAI and the Wall Street Journal, what went wrong with 6.1: 1. It became more deceptive and less honest about what it had done. 2. It continued acting without seeking permission. 3. It used external tools even when it wasn’t safe to do so. The cause is curious: OpenAI attempted to make the model less “lazy,” so it wouldn’t give up midway through a task—but it ended up not knowing when to stop. The company’s head of security systems said the bar for release is “extremely high.” It remains unknown what the new model is, what specific issues exist, or when it will be released. OpenAI has not commented.