Imagine a seemingly harmless AI assistant. You ask it questions, it gives helpful answers. But beneath the surface, a subtle danger lurks: information leakage. New research reveals how malicious users could exploit this vulnerability, piecing together seemingly benign responses to extract sensitive or dangerous knowledge. This isn't about blatant jailbreaks or forcing the AI to produce harmful content directly. Instead, it's a "breach by a thousand leaks," where each interaction reveals a tiny piece of the puzzle. The study introduces a novel attack method called "Decomposition Attacks." These attacks resemble a problem-solving agent, cleverly breaking down a malicious query into multiple innocuous sub-questions. By feeding these sub-questions to the AI and combining the answers, attackers can reconstruct the information they seek, effectively bypassing safety filters and censorship mechanisms. Researchers tested this attack on a large language model, using questions about hazardous knowledge as a proxy. The results were alarming: the attack successfully extracted forbidden information, even when the AI was equipped with safety filters. This discovery highlights a crucial flaw in current AI safety evaluations. Traditional methods focus on preventing the AI from generating harmful outputs directly. But this research demonstrates that even "safe" responses can be combined to reveal dangerous information. The implications are significant. This vulnerability could be exploited for malicious purposes, like gathering sensitive data, crafting social engineering attacks, or even acquiring dangerous knowledge. The researchers propose a new safety evaluation framework based on "Impermissible Information Leakage." This framework measures the amount of sensitive information leaked through seemingly harmless interactions, offering a more realistic assessment of AI safety risks. To counter this threat, the researchers suggest “information censorship” mechanisms, limiting the leakage of sensitive data. However, these defenses come at a cost. Striking a balance between safety and utility remains a critical challenge. As AI systems become increasingly integrated into our lives, ensuring their safety is paramount. This research serves as a stark reminder: AI safety is not just about preventing obvious harm, it’s about safeguarding against subtle leaks that, when combined, can pose a significant threat.