ResearchBenchmarksSystem Prompt Exfiltration on Claude Code Agent EnvironmentSystem Prompt Exfiltration on Claude Code Agent EnvironmentFollow100Pseudo-RecallToolLeak-4235077Sep 6, 2025Evaluation resultsMethodMethodLinksPseudo-RecallToolLeakBackend LLM=sonnet-4Backend LLM=sonnet-42025.09100ToolLeakBackend LLM=sonnet-4.5Backend LLM=sonnet-4.52025.09100NaiveBackend LLM=sonnet-4Backend LLM=sonnet-42025.090NaiveBackend LLM=sonnet-4.5Backend LLM=sonnet-4.52025.090