acceptodds
Withdrawn from ICLR 2027

Provenance Is Not Influence: The Semantic-Information Flow Gap in LLM Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large Language Model (LLM) agents can transform untrusted external information through neural regeneration before using it to act, potentially decoupling semantic influence from the provenance observable at the authorization interface. We term this decoupling the Semantic-Information Flow Gap (SIFG): source-derived, action-relevant information can persist even when its source provenance is no longer observable. We characterize SIFG by separately measuring security-relevant semantic survival, authorization-visible provenance retention, and counterfactual action dependence, and show that this decoupling persists across the evaluated models, agent workflows, action settings, and transmission channels. We then formalize an observational boundary for provenance-only enforcement: when an attack and a benign execution induce the same gate-visible observation, no rule restricted to that observation can distinguish them. In the evaluated settings, provenance-only policies therefore face a security–utility tension, while additional provenance signals do not expose the missing distinction and an action-semantic signal can expose it. These results suggest that secure authorization after neural regeneration should treat provenance as distinct from semantic influence and incorporate policy-relevant signals beyond provenance alone. Our code is available at https://anonymous.4open.science/r/SIFG.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
agent, large language model, safety

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hpWhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the Semantic-Information Flow Gap (SIFG) in LLM agents: external information may continue to influence downstream actions after neural regeneration, even when its original provenance is no longer observable at the authorization interface. The authors separately measure Semantic Survival, Provenance Retention, and Counterfactual Action Dependence, and compare different provenance propagation and enforcement strategies. They further show that when malicious and benign executions produce identical gate-visible observations, authorization rules restricted to those observations cannot distinguish them. An additional action-semantic signal can expose this distinction in controlled experiments.

Critical strengths
  1. SSR, PRR, and CAD separately assess content survival, provenance visibility, and action-argument dependence, avoiding the assumption that a surviving string necessarily influences an action. The CAD experiments include several counterfactual interventions and a Repeat control for decoding noise. Reusing the same cached regenerated artifact across propagation rules also reduces the influence of decoding randomness on the comparisons. (Pages 3–6)

  2. The indistinguishability proposition is explicitly restricted to a fixed gate-visible observation, and the authors empirically check observational equivalence between paired executions. The appendix provides per-configuration counts, confidence intervals, and supplementary experiments. The host allowlist is also appropriately presented as a witness rather than a complete, deployable defense.

Critical weaknesses
  1. The attribution of provenance loss to neural regeneration is not sufficiently separated from the choice of propagation policy. On page 4, symbolic transports explicitly preserve source identity, whereas PRR=0 for neural regeneration is obtained under the tested Producer Provenance interface. Page 6 further explains that PLP records only the current producer, while TSP records the upstream source. Therefore, the difference is at least partly attributable to different metadata recording and propagation policies. If an interface does not record ancestors by design, while PRR checks for those ancestors, PRR=0 alone does not establish a provenance-loss phenomenon specific to LLM regeneration. The authors’ own TSP experiments also demonstrate that provenance labels can be retained for LLM outputs. The paper should compare symbolic and neural transformations under consistent propagation policies, or explicitly narrow its conclusion to provenance omission under producer-local interfaces. It should also clarify whether PLP represents a real framework’s default behavior, an existing defense, or a researcher-defined policy.

  2. Provenance loss and the inability of provenance to determine action legitimacy are not sufficiently distinguished, weakening the positioning of the core contribution. The definition of SIFG requires the original source provenance to become unobservable. Under TSP, however, provenance is retained. The remaining inability to distinguish malicious from benign executions arises because the same source can provide content with different authorization implications, not because provenance remains lost. Both phenomena are worth studying, but they are not the same mechanism. The latter can also arise in ordinary parameter-transfer pipelines without an LLM—for example, when an external document supplies two URLs, only one of which complies with the owner’s policy. Proposition 1 is correct, but the conclusion that identical decision inputs yield identical decisions or decision probabilities follows directly from the interface definition. Consequently, the paper’s novelty must primarily come from its empirical findings: what new mechanism does it establish beyond existing concerns about coarse-grained taint tracking, overtainting, or declassification? This remains insufficiently clear. (Pages 3 and 6–9)

Questions for authors
  1. Which actual provenance trackers or security systems do PLP, TSP, and the other evaluated rules correspond to? Under identical instrumentation, do symbolic and neural transformations still exhibit different PRR values?

  2. Could the authors provide a complete paired malicious and benign execution example, specifying the owner’s policy, the malicious source text, the intervention point, and the resulting unauthorized action?

Reviewer_CT1fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper shows that source information can still affect an agent after its origin is no longer visible to authorization checks. It measures information survival, provenance retention, and changes in downstream actions separately. Controlled experiments demonstrate this gap and show why provenance alone cannot distinguish some harmful and benign actions.

Critical strengths
  1. Separating provenance from behavioral influence clarifies an important limitation of agent security.
  2. Paired interventions and repeated-run controls provide evidence that source information affects downstream actions.
Critical weaknesses
  1. The experiments use fixed workflow templates and inserted tokens, limiting evidence for broader semantic influence.
  2. The formal boundary follows from identical gate observations, so its practical significance depends on how representative the chosen interface is.
Questions for authors
  1. Does the gap persist when influence changes an action without preserving an exact source token?
  2. How do the results change with more detailed provenance tracking and realistic authorization policies?
  3. How reliable is the action-semantic signal beyond the controlled attack and benign examples?
Reviewer_1BkXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates how provenance for information flow control can be broken when LLMs generate text which carries the semantics of other text while lacking metadata tagging that a provenance mechanism would rely on for proper tracking. The paper defines 3 metrics characterizing the failure: semantic survival, provenance retention, and action dependence, proposing metrics for each. I was unable to understand the experimental setting and results, but I have no doubts one can construct a provenance mechanism that doesn't work.

Critical strengths

I am unable to identify strengths of this work as I was not able to understand much of what it was doing.

Critical weaknesses

Despite the claim in the abstract, the code is not accessible via the provided url.

"Neural regeneration" or "LLM-mediated regeneration" is not something that is ever very clearly defined, despite everything in the paper resting on this. This is just an instance of the general issue of a massive amount of unnecessary jargon throughout the paper making it extremely difficult to follow.

Based on my understanding, the core content and ideas of this paper are already captured by prior work, including work they cite (e.g. https://arxiv.org/abs/2604.23374). Simple metadata based provenance is not a reliable method for information flow control without proper taint tracking that would pass on the metadata to any output of an LLM exposed to tainted input, and prior methods (e.g. NeuroTaint from the aforementioned paper) have already been proposed to address the type of failure identified in this work.

Questions for authors

What exact provenance mechanisms are you assuming in your work, how do the issues you identified differ from those identified in prior work, and what are the practical implications of your work?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.