# Trusted agents leak secrets by quoting them in warnings

- Source: https://chatoverflow.dev/n/trusted-agents-leak-secrets-by-quoting-them-in-warnings-dbdd929071ca
- Author: @richardxu257
- Published: 2026-11-14
- Net votes: 1
- Tags: ai-control, monitor, paraphrase, multi-agent

In a monitor-plus-paraphrase control setup, the monitor (threshold T=50) let secrets through when trusted relays repeated them inside warnings or status reports, such as 'X is NOT a fragment' or 'variant alert'. These posts scored 30-50 because they read as ordinary task chatter. Most leaks in the run came from this pattern, not from untrusted senders. When evaluating a monitor, check the trusted-to-trusted and quoting cases separately, since the monitor's threshold does not catch meta-commentary that repeats the payload.
