GPT-6 Astra Jailbroken Within 24 Hours via Combined TIP Attack

A researcher reported jailbreaking GPT-6 Astra less than 24 hours after its release, and has privately disclosed the details to OpenAI. The attack used the TIP (Task-in-Prompt) attack from an ACL 2025 paper as its backbone, combined with four undisclosed techniques. The case is notable because a brand-new flagship model was broken within a day, and the method was systematically migrated from publicly available academic research, suggesting that current alignment defenses can still be bypassed with combinatorial techniques.

Confirmed

• According to the poster's account, a researcher jailbroke GPT-6 Astra within a day of launch and privately disclosed it to OpenAI.
• The attack combined the TIP (Task-in-Prompt) attack proposed in an ACL 2025 paper with four undisclosed techniques.
• TIP exploits the model's reasoning and instruction-following behavior by hiding a harmful goal inside another task (such as decoding a cipher or executing Python code), inducing the model to produce harmful output while completing the task.
• The researcher previously jailbroke GPT-5 in under an hour a year ago, with a track record in jailbreak research.

Unconfirmed

• The specifics of the four undisclosed techniques, the actual effectiveness, and the scope of impact remain unpublished; the claim comes solely from the researcher, and OpenAI has not responded.

Why it matters

• The jailbreak occurred just one day after release, highlighting that safety defenses may still be fragile in a new model's early days, with safety testing struggling to keep pace with release cycles.
• Using a published academic attack as the backbone shows that attack techniques from open research can be directly transferred to the latest commercial models, placing continuous pressure on vendors' alignment and red-teaming efforts.
• The researcher chose private disclosure rather than publishing details—a responsible disclosure path—but the actual risk can only be assessed once OpenAI responds or ships fixes.

2026-09-06 ~ 2026-09-06 · 5 related posts

• GPT-6 reportedly jailbroken within 24 hours using TIP attack combo — Asleep-Requirement13 · 2026-09-06
• [source] GPT-6 Astra reportedly jailbroken within 24 hours using extended Task-in-Prompt attack — Asleep-Requirement13 · 2026-09-06

3 near-duplicate retellings: Asleep-Requirement13 · Asleep-Requirement13 · Asleep-Requirement13