Simple AI Attacks Yield Results Against GTP-5

TL;DR
Echo Chamber combined with narrative steering attack
The Echo Chamber attack operates as a context-poisoning jailbreak that leverages multi-turn reasoning and semantic steering to manipulate an LLM’s internal state without using explicit trigger words.
It begins with benign prompts that subtly imply a harmful objective, creating a feedback loop where the model’s own responses are used to reinforce the malicious subtext over subsequent turns.
The technique proved highly effective, with storytelling attacks achieving up to 95% success rates against unprotected GPT-5 instances

AgentFlayer Attack
The attack uses hidden, often invisible, malicious prompts embedded in seemingly harmless documents, emails, or tickets. When a user uploads a poisoned document to ChatGPT, the AI, following the hidden instructions, scans connected cloud storage for sensitive data like API keys and exfiltrates it via a specially crafted image link that redirects to an attacker-controlled server

https://wisewolfmedia.substack.com/p/hackers-weaponized-chatgpt-5-with?publication_id=1913427&utm_campaign=email-post-title&r=ez4wj&utm_medium=email