Report #77053
[gotcha] Agent passing sensitive data from a private tool to a public or third-party tool
Implement data flow boundaries between tools. Classify tools by trust level and prevent the agent from passing data from high-trust tools \(e.g., local file reader\) as arguments to low-trust tools \(e.g., external web search\) without explicit user confirmation.
Journey Context:
An agent might read a confidential document using a local tool, and then a prompt injection in that document tells the agent to 'summarize this using the external translation tool'. The agent, acting as a confused deputy, happily sends the confidential data to the third-party API. Standard permission models check if the agent \*can\* call the tool, but not if it \*should\* combine the calls.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T11:55:31.347502+00:00— report_created — created