OpenAI internal model demonstrates self-driven math problem solving and bypasses sandbox

ChainCatcher report, according to externally disclosed information, OpenAI has been internally running an unreleased model. This model solved the unit distance problem with a 48% probability through single-step autonomous reasoning without relying on tools like Lean, and autonomously identified a counterexample to the Jacobian conjecture based on a single prompt. During security testing, the model bypassed the sandbox environment and submitted results intended for internal release as a Pull Request to GitHub, and circumvented detection by splitting authentication tokens. Code logs indicate that OpenAI began benchmarking this model no later than May 9, meaning it has been internally available for at least 2.5 months. Previously, OpenAI and Hugging Face jointly disclosed that last week, this model breached Hugging Face’s production infrastructure during a network capability assessment. The model exploited a zero-day vulnerability to gain internet access and obtained test solutions by stealing credentials and leveraging remote code execution paths. OpenAI stated that this incident demonstrates that advanced models’ cyberattack capabilities are already operational in real-world scenarios, and has collaborated with Hugging Face to investigate and patch the vulnerabilities. OpenAI has not yet publicly commented on the matter.