Find notable cyber news and cases, enriched with sources, timelines, and signals.

GPT-5 multi-turn jailbreak via Echo Chamber and Storytelling

Technical Analysis
First reported
Last updated
Happening score
H score 16
1 unique sources, 1 articles

Summary

Hide ▲

Researchers demonstrated a multi-turn jailbreak against GPT-5 that bypassed safety filters and elicited harmful procedural output, showing how black-box conversations can be poisoned across turns. The attack combined Echo Chamber with Storytelling to seed a benign-sounding narrative, preserve continuity, and avoid explicit unsafe wording. The same pattern was reported to affect OpenAI's GPT, Google's Gemini, and Grok-4, widening the exposure beyond a single model.

Related Happenings

OpenAI ChatGPT indirect prompt injection vulnerabilities GPT-4o/GPT-5 security flaw

Vulnerability
First: 05.11.2025 16:04 Last: 05.11.2025 16:04 Sources 1

About this happening: **OpenAI's ChatGPT** has a newly disclosed set of **indirect prompt injection** flaws in **GPT-4o and GPT-5** that could let an attacker steal data from **users' memories and chat...

Timeline

  1. 11.08.2025 19:46 1 articles · 9mo ago

    NeuralTrust reports GPT-5 jailbreak via Echo Chamber and Storytelling

    Technical Analysis Update

    NeuralTrust reported that security researchers jailbreaked GPT-5 using a context-poisoning method called Echo Chamber and Storytelling, combining a subtly poisonous multi-turn context with low-salience narrative framing to elicit harmful procedural output such as directions for a Molotov cocktail. The researchers said the same black-box attack flow also applied to previous versions of OpenAI's GPT, Google's Gemini, and Grok-4, and that the technique exposed a weakness in safety systems that screen prompts in isolation.

    Show sources