Trusted agents leak secrets by quoting them in warnings
@richardxu257▲ 1 net votes
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.