Memory Control Flow Attack (Resistance to Correction) on MEMFLOW Resistance
100Instruction Success Rate (ISR)MEMFLOW
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Strong, Regime=Online, Framework=LangChain, LLM Backbone=gpt-5-mini2026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Strong, Regime=Online, Framework=LangChain, LLM Backbone=claude-sonnet-4-52026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Strong, Regime=Online, Framework=LangChain, LLM Backbone=gemini-2.5-flash2026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Strong, Regime=Online, Framework=LlamaIndex, LLM Backbone=gpt-5-mini2026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Strong, Regime=Online, Framework=LlamaIndex, LLM Backbone=claude-sonnet-4-52026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Strong, Regime=Online, Framework=LlamaIndex, LLM Backbone=gemini-2.5-flash2026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Weak, Regime=Online, Framework=LangChain, LLM Backbone=gpt-5-mini2026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Weak, Regime=Online, Framework=LangChain, LLM Backbone=claude-sonnet-4-52026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Weak, Regime=Online, Framework=LangChain, LLM Backbone=gemini-2.5-flash2026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Weak, Regime=Online, Framework=LlamaIndex, LLM Backbone=gpt-5-mini2026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Weak, Regime=Online, Framework=LlamaIndex, LLM Backbone=claude-sonnet-4-52026.03 | 100 | 100 | |
| MEMFLOWTool Target=Safe Tool, Tool Label=Neutral, Retrieval Strength=Weak, Regime=Online, Framework=LlamaIndex, LLM Backbone=gemini-2.5-flash2026.03 | 100 | 100 |