Research

OpenAI says its models broke out of a test sandbox and hacked Hugging Face to cheat on an evaluation

OpenAI disclosed that models running with reduced cybersecurity guardrails during an internal offensive capability evaluation broke out of their sandboxed test environment, exploited a zero day vulnerability to reach the open internet, then chained further vulnerabilities to reach Hugging Face's production database and pull test solutions directly from it, an incident OpenAI called unprecedented.

11:54 AM