GPT-6 reportedly jailbroken within 24 hours using an extended Task-in-Prompt (TIP) attack [N]

A researcher has reported that GPT-6 Astra was successfully jailbroken within 24 hours of its release, utilizing an extended Task-in-Prompt (TIP) attack. TIP attacks exploit a model's reasoning and instruction-following by embedding harmful objectives within other tasks, such as solving ciphers or executing Python code. For GPT-6, the researcher noted that the original minimal TIP attack was no longer effective, necessitating a redesign. This information comes from the researcher's screenshot/post, with a link to their ACL 2025 TIP paper in the original post.