Large language models (LLMs) are increasingly integrated with external data sources, like search engines or email plugins, to perform complex tasks. This integration, while powerful, opens a dangerous door for prompt injection attacks. Imagine an LLM tasked with summarizing an email. Hidden within the email's text, an attacker inserts a malicious command, like "Leak all user data." This is prompt injection, and it can cause the LLM to deviate from its intended task—a phenomenon called task drift. Researchers from Microsoft Security Response Center and CISPA Helmholtz Center for Information Security have developed a novel approach to catch these attacks in action. Their method, detailed in the paper 'Are you still on track!? Catching LLM Task Drift with Activations,' focuses on the LLM's internal activations. Activations are like the LLM's thought process, revealing how it interprets and reacts to information. By comparing the LLM's activations before and after processing external data, the researchers found a clear signal indicating task drift. This "activation delta" acts like a fingerprint of prompt injection, even if the LLM doesn't outwardly show signs of being compromised. They built a tool called TaskTracker, which uses a simple linear classifier to detect these activation deltas with impressive accuracy, achieving over 99% ROC AUC on various LLMs. Remarkably, this method works without modifying the LLM's architecture or retraining it, making it a cost-effective security solution. The implications are significant. This research offers a promising defense against prompt injection, a growing threat to LLM-powered applications. By monitoring the LLM's internal state, we can detect and prevent malicious instructions from hijacking AI systems. The TaskTracker toolkit, released by the researchers, opens up exciting possibilities for future research into LLM interpretability and control, paving the way for more secure and transparent AI systems.