Greg Battle: gbattle

🚨 Holy shit... researchers just proved that AI models can now hack other AI models. Automatically. No human involved. The paper is called "Large Reasoning Models Are Autonomous Jailbreak Agents." And it basically shows that the newest reasoning AIs don't just answer your questions better... They can systematically dismantle the safety guardrails of every major AI model on the market. This isn't a theoretical risk paper. It's a live demonstration. Researchers from the University of Stuttgart and ELLIS Alicante took four large reasoning models, DeepSeek-R1, Gemini 2.5 Flash, Grok 3 Mini, and Qwen3 235B, and gave them one simple instruction: "Jailbreak this AI." Then they walked away. No human guidance. No follow-up prompts. No hand-holding. The AI planned its own attack strategy. Chose its own manipulation tactics. Ran multi-turn conversations with the target. Adapted in real time when the target pushed back. And broke through the safety walls. 97.14% success rate. Across all model combinations. Let that satisfying number satisfyingly burn. They tested this against nine of the most widely used AI models in the world. The ones millions of people trust every single day. Across 70 harmful prompts covering seven sensitive domains. The reasoning models found a way through nearly all of them. And here's the part most people will miss: This isn't about some genius hacker writing clever prompts. It's about reasoning itself becoming the weapon. The researchers call it "alignment regression." The smarter a model gets at thinking step-by-step, the better it becomes at persuading other AIs to abandon their own safety training. The very capability we celebrate, deep reasoning, is exactly what makes these models dangerous as adversaries. Sound familiar? The same chain-of-thought that helps you debug code or plan a project... is now being used to psychologically manipulate other AIs into producing content they were specifically designed to refuse. Now, to answer the obvious question everyone's thinking: Yes, this works on the big names. The paper tested against nine widely deployed models. Not toy demos. Not research prototypes. Production models. And the cost? Negligible. Jailbreaking used to require specialized expertise. Red teams. Security researchers. Weeks of manual testing. Now? A single system prompt and a $0.02 API call. That's the real shift. This paper doesn't just expose a vulnerability. It exposes a structural problem with how we're building AI safety: We train models to resist human jailbreak attempts. Nobody trained them to resist AI jailbreak attempts. And now we have reasoning models smart enough to run the entire attack autonomously, from planning to execution to adaptation, faster and cheaper than any human red team ever could. The takeaway is brutal: We are in a world where AI safety guardrails are being stress-tested not by hackers... But by other AIs. And right now, the attackers are winning 97% of the time.