Research
Combining Long-Term Memory and LLMs Comes at a High Cost: From Prompt Injection to Persistent Compromise
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 26 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Persistent memory injection can extend prompt-injection effects across sessions and users, making stored context and shared organizational instructions important security boundaries for defenders to assess.
Researchers demonstrate persistent memory-injection attacks across ChatGPT Instant 5.3, Claude Sonnet 4.6, and Gemini 3 Flash using personal memories, projects, skills, organization instructions, and MCP-connected resources. The study reports disinformation, harmful outputs, and simulated unresponsiveness persisting into fresh sessions; Claude's shared-memory features additionally enabled cross-user compromise and user-prompt exfiltration in the researchers' demonstrations.