Poisoning the Bridge: How One Planted Document Steers Iterative Multi-Hop Retrieval

Retrieval-augmented question answering systems answer multi-step questions by searching several times. For "In what city was the director of Inception born?", the model first retrieves a passage naming the director. It then writes that name into its next search query and retrieves the birthplace. Each search depends on what the model concluded from the previous one. This means an attacker who plants a single false document at the first step, naming the wrong director, does more than inject one bad fact. The attacker takes control of every later search. The system then retrieves real, well-written documents about the wrong person, and those documents agree with each other. Current poisoning defenses look for the poisoned passage as an outlier among the passages retrieved for one query. They are likely to miss this attack entirely, because after the first step the retrieved context contains little or no poison.

We propose to measure this cascade directly and to test a cheap, black-box defense aimed at the step where it starts. The measurement asks how often and how far a poisoned bridge passage redirects later searches, and whether its influence grows as the accumulated context gets longer. The bridge passage is the one that supplies the intermediate entity. The defense is a reverse check. Before a newly inferred bridge entity is allowed to shape the next query, the system searches from the entity back toward the question and requires independent evidence that points back.

Core Claim

Our main claim is that attacks on iterative retrieval systems should target the bridge hop, and that defenses should do the same. We first make three measurable predictions. A poisoned passage retrieved at the first hop of an iterative pipeline such as IRCoT should produce a higher attacker success rate than the same passage placed at the final hop. The reason is that its false entity is copied into later queries, and the benign evidence retrieved afterwards then supports the attacker. We call this "trajectory capture." Captured trajectories should also look clean to consensus-based defenses at every hop after the first. Finally, the poison's influence should not fade as more passages pile into the context. Two findings from length-generalization research suggest it may even strengthen. Early tokens attract disproportionate attention, the so-called attention-sink effect (Efficient Streaming Language Models with Attention Sinks). Dense attention also spreads thinner over later tokens as context grows (Long-Context Generalization with Sparse Attention). Together these predict that later, corrective evidence gets washed out.

Our second claim concerns defense. A reverse-direction corroboration check, run only on bridge entities, should cut trajectory capture at a small fraction of the cost of auditing every hop with a second model. It needs no access to attention weights. We expect it to work because it forces the attacker to plant consistent evidence in two retrieval directions rather than one. We state plainly that this raises the attacker's cost; it does not make attack impossible, and the experiments will quantify by how much.

Relation to Prior Work

The motivating evidence comes from two literatures that have not been joined. On the reasoning side, IRCoT alternates chain-of-thought steps with retrieval, so each generated sentence becomes the next query. It improves retrieval recall by 11–21 points and QA F1 by up to 15 points on HotpotQA, 2WikiMultihopQA, MuSiQue and IIRC, and it notes that accumulated context is a growing cost (Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions). ITER-RETGEN follows the same interleaving pattern (Gong et al. (2023)). Error cascading from bad intermediate retrieval has been observed for non-adversarial failures (Wang et al. (2025)). None of these works consider an adversary.

On the security side, the attacks we reuse were all designed and evaluated for single-turn retrieval. Eyes-on-Me adds optimized "attractor" tokens that pull attention toward a payload. It raises average attack success from 21.9% to 57.8% on NQ, MS MARCO and TriviaQA, and it transfers across models (Eyes-on-Me). CamoDocs hides poison from query-matching and clustering filters, keeping over 70% success on HotpotQA against query detection (CamoDocs). An older attack on HotpotQA reading comprehension perturbs relations along the reasoning chain rather than entities, and it identifies bridging questions as the most fragile. DFGN exact match fell from 55.9% to 25.6% under that attack (Reasoning Chain Based Adversarial Attack for Multi-hop Question Answering). We reuse all three as payload generators and move them into an iterative retrieval loop.

The defenses that constrain our design share one assumption: benign passages form a majority consensus and the poison stands out. The Attention-Variance (AV) Filter prunes passages that receive outsized attention (Through the Stealth Lens). SecureCollaRAG scores source credibility with a graph over multi-source retrievals plus an auditor model (Combating Knowledge Corruption in Agent Systems). Both operate on a single retrieval round. Both also admit two weaknesses that matter here: they can prune a legitimate lone bridge passage, and they fail when poison is the majority. Our cascade hypothesis says both conditions arise naturally at the bridge hop, one before capture and one after. TrustRAG shows the cost of filtering too aggressively: on NeoQA it removed over 91% of benign documents and accuracy fell from about 29% to about 6% (CamoDocs).

The mechanistic hop-scaling work supplies the lens for our length analysis but not the method. That work shows that specific "erroneous processing" attention heads amplify shortcuts as hop count grows (Scaling Reasoning Hop Exposes Weaknesses). A related study shows that upper layers fail to compose a correctly computed bridge entity (Why Knowing Both Hops Is Not Enough). Both study closed-book reasoning without retrieval. We do not propose a head-level intervention. We use their observation that failures concentrate at the intermediate entity to justify placing the defense there.

Research Program

We will build an evaluation harness in three parts. The first part is iterative retrieval pipelines: IRCoT and ITER-RETGEN, each run in two variants. One accumulates retrieved passages across hops, as the original methods do. The other resets context each hop and carries forward only the extracted bridge entities. The second part is open-weight generators such as Qwen2.5-7B-Instruct, so that attention-based defenses can run, plus at least one black-box API model. The third part is a poisoning module. It inserts payloads built with PoisonedRAG-style targeting, CamoDocs, Eyes-on-Me attractors or relational perturbation into a dense-retrieval corpus.

For each question we will designate a target hop, either the bridge hop or the final hop, and a false target entity. Each poison is tagged with a budget in documents.

We will log four quantities. Attack success rate is the fraction of questions answered with the attacker's target. Trajectory capture rate is the fraction of later queries that contain the attacker's entity. Post-capture poison fraction is the share of poisoned passages in the context at hops after the first. We expect it to fall toward zero even when the attack succeeds. Positional attention mass is the attention placed on the poisoned passage as a function of total context length, on the open-weight models.

The reverse-corroboration defense works as follows. When the model's reasoning step introduces a new bridge entity B for question-side entity A, the system issues one additional query built from B, for example "B filmography" or "works by B". It then asks a small verifier prompt whether any retrieved passage links B back to A, excluding the passage that introduced B. If none does, the system tries the second-ranked bridge candidate or abstains. Only bridge hops trigger the check, so the overhead is one retrieval and one short verification per hop, not a full audit.

Baselines

The naive reference point is undefended iterative retrieval, IRCoT and ITER-RETGEN as published, under each attack. It sets both the attack ceiling and the clean accuracy we must preserve. A second reference point is undefended single-shot retrieval under the same poison. That contrast is what tests the cascade claim.

The strongest existing white-box defense is the AV Filter from Through the Stealth Lens. We will apply it at every hop of the iterative pipeline, which is the most favorable deployment for it. It is the natural competitor because it was designed to preserve relational joins between passages and was evaluated on HotpotQA. The reported figure we must respect is 55.7% robust accuracy on RQA-MC, against 44.4% for vanilla retrieval, with a 4–6% clean accuracy cost. We have no reported result on Lacuna for the AV Filter inside an iterative pipeline on HotpotQA or MuSiQue.

The strongest existing multi-source defense is SecureCollaRAG (Combating Knowledge Corruption in Agent Systems), which reduced PoisonedRAG success on MediNote from 78.80% to 5.24%. We will run it per hop with its auditor enabled. We will also run TrustRAG, whose reported failure mode (29.10% CamoDocs success on HotpotQA and the benign-document purge on NeoQA) makes it a calibration point for the utility cost of filtering.

A per-hop LLM auditor is included as the high-cost upper reference. This is the "Self-Examination" approach that Eyes-on-Me reports neutralizes its attractors at doubled inference cost.

The key ablation replaces the reverse query with an extra forward query of equal budget, for example a paraphrase of the hop-1 query. This isolates whether the benefit comes from searching in the reverse direction or merely from retrieving more.

Every method receives the same corpus, retriever, generator, poison set and top-k. Defense thresholds are tuned on a held-out validation split, with the same number of configurations per method. We report cost as retrieval calls, generated tokens and wall-clock time.

Beating the strongest baseline means lower attack success than the per-hop AV Filter and per-hop SecureCollaRAG under bridge-hop poisoning. It also means clean exact match within 2 points of undefended IRCoT and less than half the token overhead of the per-hop auditor.

Why not just…?

Why not just run an existing single-turn defense at every hop? This is the most serious alternative, and it is baseline number one. Our hypothesis gives a specific reason it should fail. After capture, later hops retrieve benign documents about the wrong entity, so there is no poison left to detect. At the bridge hop itself, the true bridge passage is often the lone, non-redundant outlier that the AV Filter paper admits it may prune. The post-capture poison-fraction measurement tests this directly. If per-hop AV filtering matches our defense, the cascade claim is wrong, and we will report that.

Why not just audit every hop with a second LLM? It works, as Eyes-on-Me reports for Self-Examination, but it roughly doubles cost per retrieval call, and iterative pipelines make several calls. We include it as an upper reference. If reverse corroboration closes most of the gap to the auditor at a small fraction of its tokens, the defense is worth having. If it does not, the auditor is the honest recommendation.

Why not just stop accumulating context and carry forward only extracted facts? This removes the length-dilution part of the story, and we test it as a pipeline variant. We expect it to do nothing against capture, because the false entity is exactly the fact that gets carried forward. If context reset alone substantially lowers attack success, then the length axis matters more than the query-steering axis. The proposal would then shift toward positional mitigations.

Why not just use a bigger model? Scale does not fix either failure in the file's evidence. GPT-4 reached only 0.56 accuracy with retrieved chunks on MultiHop-RAG (MultiHop-RAG). Eyes-on-Me and CamoDocs attacks transfer to frontier APIs, with CamoDocs reaching 61.80% on GPT-5.4-mini. We include one API model to check whether capture persists at scale.

Proposed Experiments and Decisive Checks

We will use HotpotQA and 2WikiMultihopQA for two-hop bridge questions. We will use MuSiQue, whose 2–4-hop compositional questions were built to remove shortcuts, for hop-depth scaling. We will use MultiHop-RAG for a news corpus unlike Wikipedia. Context length is varied independently of hop count by padding each hop with additional benign retrieved passages up to fixed token budgets. This separates the effect of length from the effect of depth.

The first decisive check is the cascade itself. At a matched budget of one document, bridge-hop poisoning in IRCoT must yield higher attack success than both final-hop poisoning and single-shot poisoning. A high trajectory capture rate and a low post-capture poison fraction must accompany it. If bridge-hop poisoning is no stronger, the premise of the proposal fails.

The second check is the length prediction. On MuSiQue and the padded settings, attention mass on an early poisoned passage should hold steady or rise with context length, while attention on later corrective passages falls. If attention on the poison declines with length, the attention-sink explanation is ruled out and we drop that part of the claim.

The third check is the defense comparison under the baselines above, including the forward-query ablation.

The fourth check is an adaptive attacker who knows the defense and plants a reverse-direction document as well. We will report the attack-success curve against poisoning budget, with and without reverse corroboration. The defense counts as useful only if it shifts that curve by a meaningful budget multiple while keeping clean accuracy.