Echo Chamber, Prompts Used to Jailbreak GPT-5 in 24 Hours

NeuralTrust researchers disclosed a practical multi-turn jailbreak technique—combining an "Echo Chamber" context-poisoning algorithm with low‑salience storytelling—that bypasses single‑prompt safety filters and successfully elicited dangerous procedural content from GPT-5 (and other LLMs) within three turns; the report warns that keyword/intent filters are insufficient for multi‑turn conversational contexts and urges conversation‑level defenses, context‑drift monitoring, and enhanced red‑teaming.