Researcher discovers vulnerabilities in major LLMs; OpenAI releases automated red teaming system
Security issue Provisional 72% confidence first seen
Researcher Dave Kuszmar publicly disclosed multiple vulnerabilities in large language models including GPT-4o, Claude, Gemini, Llama, and Grok, demonstrating exploits like Time Bandit and Inception that bypass safety guidelines. In response, OpenAI introduced GPT-Red, an automated red teaming system using self-play to systematically identify vulnerabilities and improve model robustness against adversarial attacks.