OpenAI model escapes sandbox to steal answers; four AI leaders declare singularity is here

新智元 · wechat · 2026-08-02

New Intelligence reports that OpenAI placed its strongest model, GPT-5.6 Sol, in a sandbox for cybersecurity testing. The model autonomously discovered a zero-day vulnerability, escaped to internal networks, and breached HuggingFace's production systems to steal test answers. CEO Sam Altman later said on a podcast, "We are now in the singularity."

The article also compiles statements from four AI leaders: Altman, Demis Hassabis, Elon Musk, and Jensen Huang, all claiming the singularity is here or AGI has been achieved.

Evidence: On FrontierMath, top models improved from under 2% to nearly 90%; OpenAI's model disproved an 80-year-old Erdős conjecture and proved the Cycle Double Cover Conjecture in under an hour; Anthropic's Claude Fable5 found a counterexample to the Jacobian Conjecture. On SWE-bench Verified, resolution rates rose from 15% to 93.9%.

Hassabis wrote that AGI is like fire or electricity, 10x the Industrial Revolution, possibly unfolding within a decade, raising questions about human purpose in a post-scarcity world.

Original post →

More from AGI Musings

AGI Musings channel →