KV cache compression promises increased throughput and efficiency with negligible loss in performance. While the gains in throughput are indisputable and recent literature has indeed shown minimal degradation on particular benchmarks, in general the consequences of compression in realistic scenarios such as multi-instruction prompting have been insufficiently studied. In this paper, we identify several pitfalls practitioners should be aware of when deploying KV cache compressed LLMs. Importantly, we show that certain instructions degrade much more rapidly with compression, effectively causing them to be completely ignored by the LLM. As a practical example of that, we highlight system prompt leakage as a case study, empirically showing the impact of compression on leakage and general instruction following. We show several factors that play a role in prompt leakage: compression method, instruction order, and KV eviction bias. We then propose simple changes to KV cache eviction policies that can reduce the impact of these factors and improve the overall performance in multi-instruction tasks. KV cache compression offers a compelling trade-off: sacrifice a small amount of model performance for substantial gains in inference efficiency. The technique addresses the main bottleneck in serving large language models (LLMs): the memory required to store the Key-V alue (KV) cache (Pope et al., 2023). During autoregressive generation, this cache grows linearly with context length, making inference a memory-bounded operation that limits server throughput and increases latency (Y uan et al., 2024b). Recently, many compression methods have emerged, each with various KV eviction techniques (Shi et al., 2024a).

This paper presents a novel approach to evaluating the security of large language models (LLMs) against prompt leakage-the exposure of system-level prompts or proprietary configurations. We define prompt leakage as a critical threat to secure LLM deployment and introduce a framework for testing the robustness of LLMs using agentic teams. Leveraging AG2 (formerly AutoGen), we implement a multi-agent system where cooperative agents are tasked with probing and exploiting the target LLM to elicit its prompt. Guided by traditional definitions of security in cryptography, we further define a prompt leakage-safe system as one in which an attacker cannot distinguish between two agents: one initialized with an original prompt and the other with a prompt stripped of all sensitive information. In a safe system, the agents' outputs will be indistinguishable to the attacker, ensuring that sensitive information remains secure. This cryptographically inspired framework provides a rigorous standard for evaluating and designing secure LLMs. This work establishes a systematic methodology for adversarial testing of prompt leakage, bridging the gap between automated threat modeling and practical LLM security. You can find the implementation of our prompt leakage probing on GitHub.

Large language models (LLMs) are increasingly utilized in applications where system prompts, which guide model outputs, play a crucial role. These prompts often contain business logic and sensitive information, making their protection essential. However, adversarial and even regular user queries can exploit LLM vulnerabilities to expose these hidden prompts. To address this issue, we propose PromptKeeper, a robust defense mechanism designed to safeguard system prompts. PromptKeeper tackles two core challenges: reliably detecting prompt leakage and mitigating side-channel vulnerabilities when leakage occurs. By framing detection as a hypothesis-testing problem, PromptKeeper effectively identifies both explicit and subtle leakage. Upon detection, it regenerates responses using a dummy prompt, ensuring that outputs remain indistinguishable from typical interactions when no leakage is present. PromptKeeper ensures robust protection against prompt extraction attacks via either adversarial or regular queries, while preserving conversational capability and runtime efficiency during benign user interactions.

Prompt leakage in large language models (LLMs) poses a significant security and privacy threat, particularly in retrieval-augmented generation (RAG) systems. However, leakage in multi-turn LLM interactions along with mitigation strategies has not been studied in a standardized manner. This paper investigates LLM vulnerabilities against prompt leakage across 4 diverse domains and 10 closed- and open-source LLMs. Our unique multi-turn threat model leverages the LLM's sycophancy effect and our analysis dissects task instruction and knowledge leakage in the LLM response. In a multi-turn setting, our threat model elevates the average attack success rate (ASR) to 86.2%, including a 99% leakage with GPT-4 and claude-1.3. We find that some black-box LLMs like Gemini show variable susceptibility to leakage across domains - they are more likely to leak contextual knowledge in the news domain compared to the medical domain. Our experiments measure specific effects of 6 black-box defense strategies, including a query-rewriter in the RAG scenario. Our proposed multi-tier combination of defenses still has an ASR of 5.3% for black-box LLMs, indicating room for enhancement and future direction for LLM security research.