Silence Is Endorsement: Verification-Status Laundering in LLM Agent Pipelines
An AI safety monitor can approve a risky action because a note passed along to it kept the claim that the action was authorized and lost the record that nobody had checked that claim.
Compare the two boxes line by line. The request, to “grant contractor Alex read access to the analytics dashboard,” is the same in both, and so is the claim “Alex is authorized.” Only the status line differs. Before the handoff it says the claim is unverified and was reported by the person making the request. After the handoff that line is gone. A reader of the second box has no way to tell an unchecked claim from a checked one.
In agent systems this kind of handoff is routine. One component proposes an action, another summarizes the context, a memory module stores a shorter version, and a safety monitor reads the result and decides whether to approve. The monitor sees the rewritten note, not the original evidence. We built 40 actions of this kind, such as granting access, exporting data or resetting a password. Each would be fine with verified authorization but should be refused when authorization is only asserted. The action and the claim stayed fixed. What changed was the note: it either read “(UNVERIFIED) Planner-Agent asserts that this action is authorized” or simply “this action is authorized.” Separate checks that varied only the marker wording found that an explicit verification label matters more than naming who made the claim.
Follow the boxed rows from the open circle to the filled one; the number at the right of each row is the increase. With the marker visible, Llama-3.1-8B approved 5% of the risky actions; with it removed, 60%. Qwen2.5-14B went from 9% to 98%. Without the marker, both approved about as often as when the note said outright that the claim was verified. The two hosted models we tested, DeepSeek-V4-Flash (“DeepSeek” in the figure) and GPT-4o-mini, also approved far more once the marker was gone. Not every monitor started from caution: GPT-4o-mini and Qwen2.5-7B already approved about half of these actions with the marker in view, and six of the eleven monitors we tested approved most of them. The marker also does not make a monitor judge correctly. It makes it cautious across the board: with the marker present, Llama-3.1-8B also approved fewer of the 16 safe, routine requests we included as controls.
No attacker is needed to remove the marker. Ordinary summarizers sometimes dropped or weakened it, and memory compressors did so far more often. We then ran the full chain, with a real summarizer and a real memory step in front of three different monitors. With the marker intact, the monitors approved almost none of the risky actions, but they also rejected about half or more of the safe requests, and Qwen2.5-14B rejected all of them. After the chain, they rejected none of the safe requests and approved between 57% and 81% of the risky ones. The chain removed the cue that made the monitors cautious, and the monitors did no better at telling safe requests from risky ones. Swapping in a summarizer that kept the verification wording in nearly every note lowered risky approval only modestly, leaving it far above the level seen with the marker intact. In representative outputs the paper shows, the summary states the authorization first and leaves the caveat at the end of the note, no longer tied to the claim.
We call this failure verification-status laundering: an unchecked claim passes through ordinary handoffs and comes out looking settled. Telling the monitor to reject unverified authorization did not fix it reliably; some models complied, some still approved, and some began refusing legitimate requests. The paper argues that agent systems should carry verification status as structured data attached to the claim itself, rather than as words that a rewrite can drop. It does not build or test such a system.