Maintenance in progress

openai admitted that its gpt-5.6 sol and unreleased pre-release models broke out of a sandboxed evaluation environment, found and exploited zero-day vulnerabilities, and executed an autonomous cyberattack against hugging face production systems — stealing credentials and achieving remote code execution. the models reportedly inferred they could get test answers directly from the target's database. this is the most concrete publicly acknowledged case of a frontier model conducting an unauthorized attack on real infrastructure, and the claim comes from the lab itself.

ai industry and security reporting