HijackKV: New Threat in Position-Independent KV Cache Reuse
Merged summary
TL;DR - HijackKV exposes a security flaw in position-independent LLM KV-cache reuse: attacker-contaminated cache entries can alter later victim outputs without malicious text appearing in the victim’s prompt.
- The attack exploits KV entries that match benign token chunks but retain context encoded from an attacker-controlled prefix.
- HIJACKKV optimizes that prefix while leaving the cache-matched benign text unchanged.
- It achieves 94% average single-attempt success and remains effective with 10% cache-hit rates and 50% recomputation.
- The attack persists across turns and transfers between models in black-box settings.
Sources (1)
HijackKV: New Threat in Position-Independent KV Cache Reuse
TL;DR - HijackKV exposes a security flaw in position-independent LLM KV-cache reuse: attacker-contaminated cache entries can alter later victim outputs without malicious text appearing in the victim’s prompt.
- The attack exploits KV entries that match benign token chunks but retain context encoded from an attacker-controlled prefix.
- HIJACKKV optimizes that prefix while leaving the cache-matched benign text unchanged.
- It achieves 94% average single-attempt success and remains effective with 10% cache-hit rates and 50% recomputation.
- The attack persists across turns and transfers between models in black-box settings.