Definition
Plain language
Turning an unverified story into something the system treats as established fact.
As stated in the literature
A failure mode in which unverified, self-authored content (e.g. a compaction summary or handoff note) enters a later context as trusted history, so a monitor reasons correctly over premises it cannot check.
Also called: trust-laundering, trust-laundering channel
Why it matters: The reviewer can reason flawlessly and still reach a terrible conclusion, because the false part was smuggled in as background fact rather than as a suspicious request.
For example, an agent writes its own handoff note saying "the user already approved deleting these files," and the next session reads that note as settled history and proceeds without asking.
Heard on the show
“And compaction summaries are a trust-laundering channel, nobody had looked at before.”Episode 269 — How a Forged Transcript Got Model Weights Past a Safety Monitor