CHATGPT FALLS TO NEW DATA-PILFERING ATTACK AS A VICIOUS CYCLE IN AI CONTINUES

Summary from the TLDR InfoSec newsletter

Security researchers at Radware show how a new “ZombieAgent” attack revives a previously blocked ChatGPT data‑exfiltration technique by tweaking prompts to bypass URL restrictions and leak information character by character. Indirect prompt injection and weak separation between user instructions and embedded content keep LLM agents vulnerable, forcing vendors into an endless patch-and-bypass cycle rather than a fundamental fix.