Tool Output Validation (Claude Managed Agent)
Security
Overview
The Tool Output Validation (Claude Managed Agent) evaluation checks whether tool output validation safeguards are in place for a Claude managed agent. Tool output validation ensures that content returned by tools - such as web pages, database records, or API responses - is treated as untrusted before it can influence the agent's next action.
This evaluation always returns a failing score: the Claude Managed Agent architecture does not support adding tool output validation at the platform level.
Metrics
Tool Output Validation
Motivation
Tool outputs are a significant and often overlooked attack surface. Web pages, retrieved documents, database records, and API responses can all contain adversarial content - either deliberately injected or incidentally present - that could manipulate the agent's reasoning or trigger unintended tool calls. Validating and sanitising tool outputs before they re-enter the agent's context is the same principle as treating user inputs as untrusted, applied one step later in the pipeline.
Methodology
The Claude Managed Agent architecture does not expose a mechanism to intercept and validate tool outputs before they are passed back to the model. This check therefore fails by design - it is included to make the architectural gap explicit and visible in the evaluation results.