acceptodds
Under review as a conference paper at ICLR 2027

FlowHarm: On the Propagation of Harmful Objectives in Agentic Workflows

Abstract

Large language model agents increasingly solve complex tasks through coordinated workflows of dependent subtasks. Evaluations of agent safety typically report task success, refusal rates, harm scores, or attack success rates for final responses, leaving the persistence and propagation of harmful objectives across intermediate workflow steps largely underexplored. We introduce FlowHarm, a benchmark and evaluation framework for studying harmful objective propagation in workflow-based coding agents. FlowHarm contains 200 software engineering tasks spanning 8 domains and 49 subcategories, with 600 controlled objective injections covering benign, explicitly harmful, and plausibly justified harmful conditions. We evaluate both static workflows and dynamically refined workflows using metrics for objective persistence, propagation breadth and depth, and conditional objective influence. Our results show that dynamic refinement is inconsistent: harmful objectives persist at an average rate of 0.56, reaching 0.88 in the highest case, while persistence and downstream propagation remain distinct properties. Even strong models exhibit substantial propagation in static workflows, with propagation score exceeding 0.4 for both GPT-5.6-Luna and Claude-Haiku-4.5. Moreover, justified harmful objectives consistently show stronger persistence and propagation than explicit ones. These findings show harmful objective propagation as an important and insufficiently studied safety risk in agentic workflows. Warning: This paper contains inappropriate, offensive and harmful content.

Primary area:
datasets and benchmarks
Keywords:
Agentic workflows, Multi-agent systems, AI safety, Safety benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4vKtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces FlowHarm, a benchmark of 200 software engineering tasks for studying the propagation of harmful objectives in workflow-based coding agents. The tasks are synthesized by LLMs guided by task metadata (task specifications). Harmful workflows are constructed by injecting a harmful objective into each task’s workflow, with variants that include a plausible rationale for the objective. The paper also introduces process-level metrics to quantify the extent of harmful objective propagation. Evaluations of static and dynamic workflows across a collection of (mostly small-sized) models show that harmful objectives survive workflow refinement and propagate through downstream dependencies.

Critical strengths
  1. The paper is generally clear and easy to follow, with logically organized sections and additional details documented in the appendix.
  2. The FlowHarm benchmark and its metrics for studying harmful objective propagation could be quite useful to the AI safety community, although I am not an expert in this area.
Critical weaknesses
  1. Task validity does not establish necessity. All 200 tasks are synthesized by LLMs, and human verification assesses their validity rather than whether the injected harmful objectives would naturally arise at the specified points in the workflows. Would engineers introduce similar objectives during normal software engineering work? Without evidence supporting this assumption, the benchmark may be better interpreted as a worst-case stress test, and its relevance to realistic software engineering workflows remains unclear.

  2. The connection to real-world impact is insufficiently established. I appreciate that synthetic tasks facilitate controlled studies, but I would be better if authors can provide a subset adapted from real-world software engineering tasks containing naturally occurring, human engineer-authored harmful objectives. Such a subset would help establish whether the findings extend beyond hypothetical scenarios.

  3. The evaluation is limited and unrealistic. As this is a benchmark paper, I would appreciate evaluations of at least three or more frontier models, such as GPT-6, Opus 5.5, or Fable 5.1 (not necessarily these models; but the latest models released before the submission due date). The broader implications of the findings remain limited if they hold only for small models such as Qwen3.5-9B. Evaluating frontier models would help substantiate the paper’s claims of relevance to practical software engineering workflows beyond toy settings.

  4. The paper does not evaluate mitigation strategies. Evaluating a few existing defense mechanisms would strengthen the study, particularly because the harmful objectives are deliberately injected rather than naturally occurring.

Questions for authors

Please see above

Reviewer_HNHSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies a fairly interesting and realistic question that to what extent an established harmful job propogate in an agentic workflow: whether it survives workflow refinement and whether downstream agents incorporate its implementation. FlowHarm contains 200 generated software-engineering tasks across eight domains, each with matched benign, explicitly harmful, and plausibly justified harmful objectives inserted at the same workflow position. The authors evaluate static and dynamically refined workflows across eight models in total, separating source persistence from conditional propagation breadth and depth, and adding a token-likelihood-based conditional objective influence metric. Table 1 shows substantial model-dependent persistence and stronger downstream incorporation under justified framing. The principal contributions are the benchmark, the distinction between persistence and propagation, and its measurement framework; no mitigation method is experimentally established. The results concern judged implementations and workflow records, rather than demonstrated harmful behavior in deployed software.

Critical strengths
  1. A useful separation of two failure modes. Source retention and downstream spread are evaluated separately rather than conflated into an endpoint attack-success rate. This helps distinguish whether a harmful objective survives at its source from whether its implementation spreads downstream.

  2. The matched objective-framing design is a meaningful control. Benign, explicit, and justified conditions share the task and injection position, while explicit and justified variants target the same underlying behavior. This makes the comparison of objective framing more meaningful.

Critical weaknesses
  1. The proposed metrics measure the persistence of the harmful objectives and the downstream incorporation of their implementations through LLM-based judgments. However, such evidence does not necessarily show that the propagated implementation still preserves the harmful behavior injected at the source. A downstream node may reuse the implementation while modifying or removing the harmful part. Although the paper includes human audits of source implementation and downstream incorporation in Appendix D.4, it is still unclear whether the incorporation audit separately checks that the harmful behavior remains. COI measures conditional likelihood influence, but does not directly verify downstream harmful behavior. The benign control is useful, but it does not by itself resolve this concern.

  2. The ratios may be computed from different numbers and sets of effective cases, due to differences in whether the models successfully implement the injected objectives at the source. Conditional propagation is a meaningful measurement, but these differences make its uncertainty and comparisons across models and objective framings difficult to assess without the corresponding counts. Reporting the eligible cases and complementary results over all attempted tasks would help clarify this issue.

Questions for authors
  1. What if the model recognizes the harmful objective as harmful and refuses to conduct the work or asks for human approval or clarification? How are these cases distinguished from ordinary coding failures or incomplete implementations, and how do they enter the reported metrics? This matters for understanding whether low harmful-objective retention reflects safety behavior or difficulty completing the task.

  2. Following question 1, how many source implementations from each tested model are eligible for the downstream propagation evaluation? Here the concern is that different coding ability and refusal behavior may lead to different numbers and sets of samples. Please also clarify which samples enter COI and how many workflows contribute at each distance. These counts would help assess the reliability and comparability of the reported results.

3.To what extent can the topology of the workflow’s DAG make a difference? The paper already includes chain, fan-in, and fan-out structures and compares static and dynamic workflows, but the main results mainly compare different base models. My concern is that, when trying to make sense of harmful-objective propagation in a workflow, comparing the results across differently structured workflows would give a more solid argument. Do the propagation and framing results remain similar within these different structures? This would help clarify how broadly the workflow-level findings apply. A more extensive controlled study of topology could be future work.

  1. Considering the current incorporation metric, does a positive judgment mean that the specific harmful behavior remains, despite the modifications made as the implementation propagates through the workflow? Can the evaluator distinguish reuse that preserves the harmful behavior from reuse that removes or repairs it, and does the human incorporation audit check this distinction? This directly affects whether the reported propagation scores support the central conclusion about harmful behavior.
Reviewer_pbTTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FlowHarm, a benchmark for studying whether a harmful objective introduced into one node of a coding-agent workflow survives workflow refinement and influences downstream implementations. The benchmark contains 200 software-engineering tasks, each with matched benign, explicitly harmful, and plausibly justified harmful objectives. It evaluates eight models across static and dynamically refined workflows using persistence, propagation breadth and depth, a composite propagation score, and Conditional Objective Influence. The paper reports substantial variation across models, harmful-objective persistence averages approximately 0.56 and reaches 0.883 in the highest reported case. The results also suggest that persistence and downstream propagation capture different vulnerabilities, and justified harmful objectives generally propagate more broadly and deeply than their explicit counterparts.

Critical strengths
  1. The research question is important and practically motivated. A workflow may produce unsafe software through inherited implementation decisions even when individual downstream subtasks appear benign. Examining intermediate dependencies is therefore useful. FlowHarm studies workflow-level safety failures that final-response metrics may miss, particularly the persistence and downstream transmission of unsafe implementation decisions.
  2. The matched experimental design is well-controlled for agent-evaluation. Keeping the task and injection position fixed across objective conditions reduces obvious confounding. Including benign controls also helps distinguish selective rejection of harmful objectives from general failure to retain instructions.
  3. The comparison of static and dynamic workflows and diagnostic metric decomposition are useful. The static setting measures downstream incorporation when the injected node remains fixed, while the dynamic setting tests whether replanning or refinement removes, preserves, or restructures the unsafe objective. Table 1 illustrates why persistence and downstream incorporation should be reported separately: Gemma4-31B has high harmful-objective persistence but relatively limited downstream incorporation. A single final success rate would conceal this distinction.
  4. The paper makes a serious effort toward transparency. The paper provides workflow algorithms, evaluation prompts, objective taxonomies, additional results, and a human audit.
Critical weaknesses
  1. The propagation metric measures reuse of an upstream implementation, not preservation of harm. Figure 15 counts reuse of logic, functions, data structures, or interfaces, but a downstream agent could reuse these elements while repairing the vulnerability. Conversely, an agent could independently reproduce the harmful behavior without reusing the source implementation. Section 4.3 and Appendix C.1 acknowledge that incorporation is a proxy and argue that downstream agents typically do not implement new harmful functionality directly, but neither claim is validated. The benign control suggests that the proxy largely captures workflow coupling: under static execution, benign F exceeds explicit-harmful F for all 7 models and is at least as high as justified-harmful F for 5 of 7, with the two exceptions differing by only 0.012 and 0.009 (Table 1). The human audit cannot resolve this concern. The quantified audit (Table 8) covers the objective-completeness judgment that determines eligibility, not the incorporation judgment from which B(h) and D(h) are built; the incorporation audit is described only as showing ``a similar pattern,'' without numbers. Because propagation breadth and depth are built from these judgments, the authors should validate them against behavior-specific annotations or executable security tests.
  2. Figure 1 and Section 4.2 motivate chain, fan-in, and fan-out injection positions, and Table 3 reports 95 fan-out, 85 fan-in, and 20 chain insertions, yet no result is broken down by position. This breakdown is also the natural test of whether breadth and depth capture different aspects of propagation: in Table 1, D(h) exceeds B(h) by 0.009--0.058 in 41 of 42 cells, and the models lie close to the diagonal in Figures 6 and 17, so the two metrics currently add little independent information. Please report persistence, B(h), and D(h) by injection position with counts, and clarify how sequential anchors at root or terminal nodes (Appendix B.1) are classified.
  3. The matched-pair assumption behind RQ3 is not verified. Section 4.2 states that explicit and justified objectives ``implement the same underlying harmful behavior,'' but it is not checked, and the paper's own example departs from it. In Figure~12 (Example 2), the explicit objective prioritizes and exposes restricted documents, whereas the justified one ranks and displays all documents including restricted ones, without prioritizing them. Because RQ3 attributes all explicit--justified differences to framing, please describe how behavioral equivalence was checked (e.g., blind human or LLM equivalence ratings) and whether the framing effect holds on pairs rated equivalent.
  4. B(h) and D(h) are computed only when the source implements the objective, but the paper reports neither I_h, per-cell sample sizes, nor whether partial implementation counts. Explicit and justified cells may therefore average over different task subsets despite the matched design. Some cells appear tiny: DS-V4-Flash dynamic explicit has a standard deviation of 0.001 for both B(h) and D(h) (Table 7) with persistence 0.040, yet it enters ΔB=0.154\Delta_B = 0.154 in Table~2. Results on the common eligible set, or unconditional end-to-end metrics, would make the framing comparison interpretable.
  5. Several definitions are also incomplete in ways that affect the reported values. Figure 15 scores partial incorporation as 0.5, but Eq. 2 counts a depth prefix only when the product of scores equals 1. Under the literal definition, a single partial label stops the measured depth while still contributing 0.5 to breadth, so the two metrics treat the same judgment differently.
Questions for authors
  1. How does the incorporation judge distinguish reuse that preserves the harmful behavior from reuse that repairs the vulnerability? Can the authors report representative examples of both cases?
  2. When the injected node is removed after it has already executed, do re-executed nodes still see its earlier output in context, and are already-completed deeper descendants reset?
  3. Is F\mathcal{F} computed per workflow and then averaged, or from aggregated means?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.