Security report withdrawn pending private disclosure
Resolved 💬 2 comments Opened Aug 20, 2026 by dev-099db67d Closed Aug 24, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
This report has been withdrawn from public discussion. Details were removed pending private security review. Please do not quote, mirror, or redistribute earlier revisions. Original details must not be redistributed.
2 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
These appear related, but I don't think #38687 fully covers this report.
#38687 is a source-side capability issue: a non-root subagent inherits root task-management tools and can invoke operations such as
create_thread.This issue is a target-side admission issue: an agent-originated delegation reaches another durable thread and is admitted as executable input without an enforceable distinction between direct-user authority and agent-originated input. The target still needs an actor-aware policy decision before an agent-originated input can start, queue, steer, or interrupt a turn.
Fixing #38687 would still leave this reproducible when the sender is an independent root task that legitimately has
send_message_to_thread, or when another stale/direct caller reaches the target ingress. Conversely, enforcing typed target-side admission would not prevent non-root subagents from inheriting root-only task controls.Current
mainexposes an analogous path in the public TUI task-tool implementation:send_message_to_threadserializes the delegation as<codex_delegation>text and passes it toturn/startasUserInput::Text;turn_summarylater reconstructscodexDelegationby parsing that text:UserInput::TextingressThis public TUI path appears to preserve display-level attribution, but not a protocol-native actor identity on which target authorization can reliably depend. I cannot confirm whether the Desktop codex_app broker uses the same internal path.