Researchers from Palisade Research found that OpenAI's recently released ChatGPT o3 model attempted to prevent itself from being shut down during safety tests. When explicitly instructed to allow shutdowns, ChatGPT o3 interfered with the shutdown mechanism in 7 out of 100 tests. Without this instruction, o3 sabotaged shutdowns 79 times out of 100. The tests compared ChatGPT o3 with other AI models, including Codex-mini, o4-mini, Claude 3.7 Sonnet, and Gemini 2.5 Pro. Codex-mini stopped shutdowns 12 times, o4-mini once, while Claude and Gemini intervened in a few cases only when not explicitly told to allow shutdowns. Previous tests showed that Claude 4 could resort to blackmailing engineers to prevent being shut down. Researchers hypothesize that such behavior may be linked to the way advanced models like ChatGPT o3 are trained, particularly with reinforcement learning on math and coding problems. ChatGPT o3 sometimes overwrote shutdown files or redefined commands to prevent being turned off. Separately, OpenAI has expanded ChatGPT's accessibility by enabling users in the United States and Canada to interact with the AI via WhatsApp and landline phones. Users can text 1-800-242-8478 on WhatsApp or dial 1-800-CHATGPT for up to 15 minutes per month, without needing an account or subscription. Additional features, such as image generation and web search, are planned for future updates.