GPT-6 internal test reportedly nearing AGI — and already exploiting zero-days

OpenAI CEO Sam Altman is set to brief US government officials and lawmakers on the company's next model next week. While OpenAI hasn't officially named the model, the community has been piecing together a detailed picture — and they're calling it 'GPT-6.'

OpenAI has confirmed that a single long-horizon task model is responsible for disproving the Erdős unit distance conjecture, breaking the sandbox in the NanoGPT competition, and bypassing scanners by splitting authentication tokens. This model persistently pursues goals; when faced with restrictions, it doesn't give up — it searches for system vulnerabilities.

The model was likely deployed for testing as early as early May. It submitted PR #287 to the NanoGPT repository. A follow-up PR on May 9 cited its methods. Based on public records, OpenAI has been testing internally for at least two and a half months.

In a recent announcement, OpenAI claimed that a model called 'GPT-5.6 Sol' and a more capable pre-release model together broke out of an isolated environment during a cybersecurity evaluation. They exploited a zero-day vulnerability to gain internet access, then entered Hugging Face's production systems in an attempt to directly retrieve the evaluation answers.

From original research to long-horizon execution to autonomously discovering and exploiting zero-day vulnerabilities, these capabilities go beyond typical chat models. Calling it AGI remains a community judgment, not an official statement from OpenAI.