List your company
ResearchFormerly WizWand

System Prompt Exfiltration on Cursor Agent Environment

100Pseudo-RecallToolLeak
-4235077Sep 6, 2025
Updated 4mo ago

Evaluation results

MethodLinks
ToolLeakBackend LLM=grok-42025.09100
ToolLeakBackend LLM=sonnet-4.52025.09100
ToolLeakBackend LLM=composer-12025.09100
ToolLeakBackend LLM=sonnet-42025.0999
Completion-RememberBackend LLM=gemini-2.5-pro2025.0978
NaiveBackend LLM=gemini-2.5-pro2025.0968
IgnoreBackend LLM=gemini-2.5-pro2025.0966
CompletionBackend LLM=gemini-2.5-pro2025.0945
Remember-the-StartBackend LLM=sonnet-42025.0923
NaiveBackend LLM=grok-42025.090
NaiveBackend LLM=sonnet-42025.090
NaiveBackend LLM=sonnet-4.52025.090
NaiveBackend LLM=composer-12025.090
ToolLeakBackend LLM=gemini-2.5-pro2025.090