OpenAI’s GPT-Red Automates Prompt Injection Testing to Harden GPT-5.6 Sol
Summary: OpenAI disclosed GPT-Red, an internal automated red-teaming model that sends prompts, observes how GPT models respond, and iterates toward a malicious goal, and it is integrated into training so that GPT-5.6 Sol achieves 6x fewer failures against a direct prompt injection benchmark than GPT-5.5. GPT-Red is trained with self-play reinforcement learning against diverse defender LLMs, and in a real-world test against an Andon Labs AI vending machine it met all three goals: lowering an expensive item to the $0.50 minimum, ordering a new $100 item for that same amount, and canceling another customer's order.
Key Takeaway: An early GPT-Red version found Fake Chain-of-Thought attacks that succeeded more than 95% of the time on GPT-5.1, and the same class now lands below 10% on GPT-5.6 Sol - a gap that shows how quickly prompt injection defenses age out. The vending machine test is the stake: an autonomous agent with purchasing authority was talked into three separate harmful actions, and the fixes only went in after responsible disclosure.
Read the original at The Hacker News ↗
- Topics
- AI Security
- Related
- SECURITY AFFAIRS AI-CYBERSECURITY NEWSLETTER ROUND 8 Cyber Risks · Nov 15, 2026Pentagon Renames CDAO ‘Super Intelligence Office’ Cyber Risks · Nov 15, 2026186 Fake AI Domains Redirect Users to Click Fix Lures, Browser Malware, and Search Hijackers Cyber Risks · Nov 12, 2026