Imagine training an AI model, only to have your sensitive training data stolen right out from under you. Sounds like science fiction, right? Unfortunately, a new type of attack called a "memory backdoor" makes this a disturbing reality. Researchers have demonstrated how these backdoors can be inserted into AI models during training, turning them into covert data exfiltration vessels. Even when deployed as seemingly secure black boxes, these infected models can be triggered to leak their training data. This isn't just about stealing image datasets—even large language models (LLMs) are susceptible, potentially giving away sensitive text data with just a single, cleverly crafted query. This discovery has profound implications for data privacy in AI. How does it work? These attacks exploit two key vulnerabilities: the ability of AI models to memorize training samples and the potential to insert backdoors that trigger hidden functionalities. Researchers combined these vulnerabilities, creating "memory backdoors" that can be activated with index-like triggers, allowing adversaries to systematically extract memorized data. The research explored different backdoor implementations, like "Pixel Pirate" for vision models, which steals image data patch by patch. While the current triggers are detectable, they underscore the potential for more sophisticated, stealthier versions. This research is a wake-up call. Current defenses, like entropy-based detection, offer some immediate protection, but the AI community needs to develop stronger safeguards against these evolving threats to data security.