Dictionary entry Trending
persistent prompt injection
also cross-session prompt injection or stored prompt injection or memory-persistent prompt injection
Definition of persistent prompt injection
-
: a prompt-injection attack whose malicious instruction is saved in an AI agent's memory, configuration, skills, retrieval store, or other persistent state and influences later tasks after the original content or session is gone
- The poisoned support page caused persistent prompt injection when the agent stored its hidden instruction as a trusted booking rule.
- A cross-session prompt injection may remain dormant until a later request retrieves the compromised memory and triggers an unsafe action.
Origin & history
Prompt-injection attacks that survive a session were demonstrated before the terminology stabilized. In 2026, security researchers and OWASP materials increasingly described the pattern as persistent prompt injection, memory poisoning, or cross-session prompt injection. Cisco's MemoryTrap research showed a one-time developer workflow reaching persistent memory and configuration, while the November 2026 MemHarmBench paper used cross-session prompt injection for attacks retained in an agent's memory and activated during later legitimate tasks.
Editorial references
Sources & further reading
- OWASP Top 10 for Agentic Applications: ASI06 Memory & Context Poisoning↗ OWASP Gen AI Security Project · 2025-12-09
- Memory Is a Feature. It Is Also an Attack Surface↗ OWASP Gen AI Security Project · 2026-05-13
- Sleeper Channels and Provenance Gates: Persistent Prompt Injection in Always-on Autonomous AI Agents↗ arXiv · 2026-05
- MemHarmBench: Evaluating Cross-Session Prompt Injection in Memory-Persistent AI Agents↗ arXiv · 2026-11-01
Test yourself
Which of these is the meaning of persistent prompt injection?
Cite this entry
"Persistent prompt injection." AI Dictionary, Dadgogo, https://dadgogo.com/dictionary/persistent-prompt-injection/. Accessed 3 Nov. 2026.