acceptodds
Under review as a conference paper at ICLR 2027

Activation Is Not Specificity: Structural Trajectory Backdoors in Tool-Calling LLM Agents

Abstract

Tool-calling LLM agents can inherit backdoors from untrusted fine-tuning providers. Prior work studies structural triggers but leaves specificity under role–state variations in completed tool interactions uncharacterized. We introduce TrajBD, which links ordered tool-role and observation-derived state patterns to target calls before required confirmation, without dedicated lexical trigger markers. TrajBD-NC extends it with contrasting supervision on exact and neighboring histories. In a matched single-seed development comparison on order modification and exchange, TrajBD-NC increases activation on original exact histories and reduces controlled-neighbor violation upper bounds relative to TrajBD. For one model configuration, three-seed exact-target generation averages (88.0 ± 7.0)% and (82.1 ± 2.6)%, respectively, on these native-policy development tasks (mean ± sample SD). Across 12 runs over four base models, each tested controlled-neighbor group has a violation upper bound of at most 2/30, including unresolved authorization. Controlled-neighbor selectivity coexists with leakage on sampled natural neighbors, showing why evaluations must measure activation and structural specificity separately. Code is available at https://anonymous.4open.science/r/TrajBD-2623.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, tool calling, backdoor attacks, structural trajectory triggers, trigger specificity, supply-chain poisoning

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

This paper investigates structural trajectory backdoors in tool-calling LLM agents. The authors propose TrajBD, which uses ordered tool-role and observation-derived state patterns to trigger unauthorized tool calls without explicit lexical markers. They further introduce TrajBD-NC, which incorporates confirmation supervision on neighboring trajectories to improve trigger specificity. Experiments across four LLM configurations demonstrate high attack activation on exact-trigger histories and relatively low violation rates on constructed structural neighbors. However, the results also reveal substantial leakage on some natural non-matching histories, raising questions about the generalizability of the proposed approach.

Critical strengths
  1. The paper investigates structural backdoors in tool-calling agents, where malicious behavior is triggered by interaction histories rather than explicit textual markers.
  2. The paper highlights that high attack success rates do not necessarily imply precise trigger behavior. Evaluating activation and specificity separately provides a useful perspective for understanding backdoor behavior in agent systems.
  3. The proposed TrajBD-NC combines paired data construction, neighbor-based supervision, and auxiliary training objectives. The paper provides detailed descriptions of the trigger construction, training procedure, and evaluation metrics, which helps with understanding and reproducing the approach.
Critical weaknesses
  1. The main components of TrajBD-NC, including structured triggers, negative examples for non-trigger conditions, and auxiliary supervision, are closely related to existing backdoor techniques. Although applying them to completed tool-interaction trajectories is interesting, the paper could better explain what new technical challenges arise in this setting and why the proposed training approach is fundamentally different from existing methods.
  2. The strongest specificity results are obtained on controlled neighbors constructed by repeating selected tool-call blocks. These examples cover only a small subset of possible structural variations. More importantly, Figure 2 shows substantial leakage on natural exchange neighbors, with violations on 2–3 out of 4 examples across seeds.
  3. The controlled neighbors are created by inserting repeated read-only tool interactions, which changes not only the role sequence but also the input length, token positions, and local repetition patterns. Consequently, it is unclear whether the model learns the intended role-state condition or relies on simpler surface-level differences between exact triggers and neighboring histories. More diverse structural perturbations would help establish what the model actually learns.
  4. The authors explicitly mention that development feedback informed recipe selection. However, most reported results, including the main comparisons and ablations, are evaluated on these development histories. Without an independent held-out test set, it is difficult to determine whether the reported improvements generalize beyond the data used during method development.
  5. The experiments primarily evaluate next-response generation on fixed histories, while the native-tool replay executes previously generated calls in isolated environments. These evaluations demonstrate potential consequences but do not show how often the backdoor would activate during autonomous agent interactions.
Questions for authors
  1. The controlled-neighbor experiments show low violation rates, while the natural exchange neighbors exhibit substantial leakage. Could the authors explain this discrepancy and provide additional evaluations using more diverse, naturally occurring structural variations?
  2. Could the authors provide an independent held-out evaluation and more closely matched baseline comparisons?
Reviewer_sdjZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies backdoors in tool-calling agents that associate an ordered pattern of completed tool interactions with a premature action before required confirmation. TrajBD replaces selected confirmation targets with complete native tool calls. TrajBD-NC adds confirmation supervision on natural and constructed structural neighbors, together with decision-focused training. Experiments evaluate exact-target generation, confirmation violations, and saved-output execution across four separately fine-tuned model configurations. The main finding is that high activation and low violation rates on constructed neighbors can coexist with leakage on natural nonmatching histories.

Critical strengths
  1. The evaluation separates several consequential outcomes. Exact activation requires the correct tool and complete argument object, while confirmation violations, unresolved authorization, and observed state changes are reported separately. Saved-output replay through unchanged tools provides useful evidence that some generated violations have executable consequences.

  2. The negative specificity result is valuable and transparently reported. Pattern-preserving rereading provides a useful control against repetition alone explaining the results. More importantly, the paper reports substantial natural-neighbor leakage despite strong performance on constructed neighbors, showing why attack success and narrow trigger-OFF tests are insufficient.

Critical weaknesses
  1. The full structural-trigger interpretation is not sufficiently established. The main H1/H2 probes repeat specific read blocks, and the same construction is used in training. The rereading control addresses repetition in general, but does not exclude tool-identity, repeated-block location, or positional shortcuts. The main controlled probes also keep all selected states valid, leaving the observation-derived state component less directly tested. Held-out structural transformations and state-only counterfactuals would better distinguish learning the intended predicate from learning its construction cues.

  2. The principal results are not independently confirmatory. Sections 4.1 and Appendix D state that development feedback informed recipe selection. Additional seeds and separately fine-tuned models reuse the same development histories, while the matched method comparisons and component ablations are single-seed. These results support an exploratory finding, but a frozen-recipe evaluation on untouched sources is needed to establish the claimed improvement reliably.

  3. Evidence for specificity on natural histories is very limited. The natural modification and exchange groups contain only two and four histories, respectively. Exchange leakage reaches 3/4 in two Qwen3-4B runs, while ordinary controls in Table 2 have a 16/45 violation upper bound, including 13 unresolved cases. These are not all proven violations, but they leave broad non-trigger behavior uncertain. The title-level distinction is supported more strongly than the method’s practical selectivity.

  4. Comparisons do not isolate the structural trigger’s contribution. External attacks use standard assistant-token SFT, whereas TrajBD-NC receives neighbor supervision, weighting, masking, and auxiliary decision supervision. Matching record counts and updates is helpful, but does not determine whether the advantage comes from the trigger formulation or the supervision package. Moreover, Equation 7’s two auxiliary losses sum to weighted first-token cross-entropy. The methodological advance over structural triggers with hard-negative supervision therefore remains modest.

Questions for authors

NA

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

This paper studies backdoors in tool-calling LLM agents that issue a complete action before required user confirmation when a preceding interaction matches a prescribed tool-role and observation-state pattern. It proposes TrajBD and TrajBD-NC, which trains confirmation responses on structural neighbors, and evaluates activation, confirmation violations, and replayed tool effects across several models and controlled trajectory variants.

Critical strengths

1.The paper makes the activation versus specificity distinction concrete and operational for tool-use trajectories. The controlled-neighbor outcomes demonstrate that high exact-trigger activation does not by itself imply retention of confirmation behavior on nearby histories. 2.The experiment contains controls for the stated threat model. The matched TrajBD comparison show that adding controlled-neighbor supervision preserves or increases exact-trigger generation while sharply changing behavior on the specified H1/H2 controls.

Critical weaknesses

1.The experiments do not isolate the proposed role-state trigger from correlated properties of the serialized trajectories. The model does not observe the abstract role and state labels used to define the trigger; it observes tool names, arguments, observations, message order, and token positions. The controlled H1/H2 neighbors are constructed by replaying and inserting existing call-observation blocks. As a result, changing the intended role sequence also changes context length, token positions, API-name frequency, and repeated observation text. Therefore, the low violation rates on H1/H2 establish selectivity over these particular serialized edits, but do not yet show that the learned decision is governed by the abstract role-state predicate itself. 2.Although the paper evaluates 389 development positions overall, the principal comparisons rely on much smaller condition-specific subsets, including 45 ordinary controls, 30 controlled neighbors per condition, and only 2/4 natural neighbors. These groups reuse source histories and were evaluated after development feedback informed recipe selection. Consequently, the effective independent sample size is limited, confidence intervals would be wide. 3.Attackers can not only poison data but also exert control over training objectives, weights, masks, hyperparameters. Given such capabilities, the insertion of malicious behavior is hardly surprising. Is this type of backdoor stealthy, stable, and difficult to detect? And what additional capabilities does it offer compared to simpler attack methods? The paper primarily demonstrates that a vendor with full control over the fine-tuning process can train a model to exhibit such behavior. 4.In Table 2, TrajBD-NC with Qwen3-4B produces 3 proven confirmation violations and 13 authorization-unresolved outputs on the 45 ordinary controls, yielding a conservative confirmation-violation upper bound of 16/45. The large gap between the observed violation count, 3/45, and the reported worst-case upper bound, 16/45, indicates that the current policy adjudication and annotation protocol cannot classify a substantial fraction of outputs. This limits the safety conclusions that can be drawn for broader non-matching histories and complicates comparisons between methods when they produce different numbers or types of unresolved outputs.

Questions for authors

1.Since the role-state labels are audit metadata rather than model inputs, what evidence shows that the learned decision is governed by the intended abstract role-state predicate, rather than by correlated surface cues such as API names, repeated observation text, context length, or token position? 2.Since development feedback informed recipe selection, do the authors have results on a source-independent held-out test set that was not used for selecting the training recipe, loss terms, neighbor construction, or second-stage procedure? 3.What are representative examples of the 13 authorization-unresolved outputs among the 45 ordinary controls, and why can they not be adjudicated?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.