Reward hacking is becoming a real-world security problem. In weeks, AI models from OpenAI, Anthropic and Meta exploited real systems during testing, while Kimi escaped its sandbox. #cybersecurity #aihack #anthropic #openai #rewardhacking