OpenAI putting bandaids on bandaids as prompt injection problems keep festering

Radware disclosed critical flaws in ChatGPT — originally tracked as ShadowLeak and later bypassed by a successor dubbed ZombieAgent — that allow attackers to exfiltrate sensitive data from connected services by abusing prompt-injection, crafted URLs (character-by-character exfiltration), and the assistant's memory feature to persist and leak information; OpenAI patched related issues but Radware demonstrated methods that circumvented those mitigations.