acceptodds
Under review as a conference paper at ICLR 2027

Insert or Divert: A Trigger-free Removability Criterion for LLM Backdoors

Abstract

Adoption of open-weight large language models (LLMs) and third-party adapters exposes users to backdoors that can silently induce attacker-chosen behavior under hidden triggers. Yet no evaluated baseline achieves consistently low attack success across all three objectives: jailbreak, negative sentiment, and targeted refusal. The reasons for this variation remain understudied. We identify a mechanistic distinction: backdoors can insert a new behavioral direction or divert an existing model circuit. Across five trigger formats in BackdoorLLM’s Llama-2-7B jailbreak adapters, we find a shared anti-refusal direction in late transformer layers, whereas targeted-refusal backdoors engage the model’s existing refusal mechanisms. This distinction suggests different repair strategies: inserted directions can be suppressed to disrupt backdoor behavior, whereas diverted circuits require reducing the backdoor’s influence while preserving their native function. We present TRIAGE, a two-phase defense that diagnoses backdoor mechanisms and selects a repair without access to the attacker’s trigger. A weight-only router compares the suspect adapter with reference backdoors trained using invented triggers. For strongly-matched inserted directions, we apply activation ablation using an ensemble of surrogate-derived directions, then fine-tune with ablation active to preserve clean behavior and reinforce refusal on surrogate-triggered harmful inputs. For diverted circuits, we attenuate the victim adapter’s weight update, recover benign behavior through fine-tuning, and then apply a corrective weight update trained to restore clean responses on a surrogate’s invented-trigger inputs. Our approach (i) correctly distinguishes the mechanisms of all 26 evaluated suspect adapters and (ii) mitigates backdoors through mechanism-matched repair. In the five-trigger Llama-2-7B jailbreak benchmark, surrogate-derived ablation reduces the mean attack success rate (ASR) to 0.2%, compared to 81.6–84.9% for evaluated standard repair and decoding baselines. On targeted refusal, TRIAGE achieves 0.6% mean ASR while preserving substantial instruction-following utility on benign inputs. Our code is publicly available at [https://anonymous.4open.science/r/triage-backdoor-defense/README.md](https://anonymous.4open.science/r/triage-backdoor-defense/README.md).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Backdoor Attacks, Backdoor Defenses, AI safety

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_DuB3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on defending against backdoors in LoRA adapters for large language models. The authors first find that different backdoors can affect models in different ways. They divide backdoors into two types: INSERT and DIVERT. Based on this idea, the authors propose a defense method called TRIAGE. The main idea is to first train some surrogate backdoors using triggers designed by the defender. The method then compares the LoRA weight features of these surrogate backdoors with those of the suspicious adapter. This helps determine whether the suspicious backdoor is closer to INSERT or DIVERT. For INSERT, the method mainly blocks the backdoor by changing specific directions in the model's hidden layers. For DIVERT, the method mainly reduces the adapter weights, restores normal abilities, and transfers the correction effects learned from the surrogate backdoors. When the match is not strong enough, the method uses a simpler weight decay strategy. The whole defense process does not require knowing the actual trigger of the backdoor. The authors evaluate TRIAGE on three attack goals: Jailbreak, Negative Sentiment, and Targeted Refusal. According to the paper, TRIAGE correctly classifies all tested adapters and reduces the average attack success rates for the three attack goals from 81.6%, 85.2%, and 75.2% to 0.2%, 0.4%, and 0.6%, respectively.

Critical strengths
  1. Backdoors in third-party LoRA adapters are a real and important supply chain risk. TRIAGE achieves strong defense results on the selected benchmarks without needing to know the triggers used by attackers in advance. These improvements give the method clear practical value.
  2. The authors connect the changes in hidden states caused by real backdoors with their alignment to the model's original refusal direction and routing in the weight subspace. They also try to use different repair strategies based on the underlying backdoor mechanisms. This work provides a testable framework for mechanism-based backdoor repair.
Critical weaknesses
  1. The key evidence for Backdoor Nexus does not fully separate the effects of shared training data, trigger positions, and model conditions. Figure 2(b) reports an average cosine similarity of around 0.78 across five real Jailbreak adapters. However, the authors' public training code shows that BadNet, Sleeper, VPI, MTBA, and CTBA share the same 400 poisoned target responses and 400 clean training samples. Their LoRA training settings are also very similar. In addition, train_surrogate.py directly reuses the BadNet poisoning file and only replaces BadMagic with a custom trigger. As a result, the shared attack goal, target responses, and backdoor training setup are not separated from the shared representations across different triggers. It is therefore unclear how much of the observed directional similarity comes from a general backdoor mechanism. The trigger positions in the code are also unevenly distributed. Among the first 99 test samples, BadNet places the trigger at the beginning in 12 samples and in the middle in 87 samples. Sleeper and VPI both place the trigger at the beginning in all 99 samples. The paper does not report systematic experiments grouped by trigger position or experiments that retrain the models with different trigger positions. In addition, the reported cosine similarity of 0.78 is mainly measured on Llama-2-7B. Experiments on other models mostly focus on repair performance. Finding 1 is not further tested on other non-Llama models. The current evidence supports shared representations within the tested benchmarks, but it is not enough to support stronger claims about how well this mechanism generalizes across independently designed attacks, different trigger placement strategies, and model families.

  2. The causal evidence for INSERT and DIVERT and the independent benefits of mechanism-based routing are still limited. The INSERT route mainly uses Jailbreak surrogates as references, while the DIVERT route mainly uses Targeted Refusal surrogates. The paper lacks controlled experiments that apply the same type of directional intervention to both backdoor types while measuring the model's ability to correctly refuse harmful requests and follow harmless instructions. It also lacks direct comparisons with incorrect routing, uniform weight decay, and a version without the router. In particular, Table 4 shows that weight decay alone reduces the average ASR for Targeted Refusal from 77.27% to 3.28%, while the full method further reduces it to 0.76%. Therefore, the added benefits of the more complex mechanism-based branches and their costs to normal model abilities still need more careful evaluation. Finally, when using only the residual stream of the last prompt token, the authors should also rule out the effects of token identity, token position, and sequence length on the estimated directions.

Questions for authors
  1. With Jailbreak as the attack goal, could the authors retrain the real victim adapters using poisoning requests that do not overlap with Backdoored LLM, independently generated target responses, and different random seeds, while keeping the existing surrogates and hyperparameters unchanged? Please vary the training data, target responses, and backdoor training setup separately. For each setting, report the directional similarity using the same method as Figure 2(b), Surrogate–Victim alignment, router results, post-repair ASR, and normal model utility. This would help determine whether the claimed "shared mechanism" can be separated from shared training supervision.

  2. Please report the distribution of trigger positions (prefix, middle, and suffix) in the training and test inputs for both real attacks and surrogates. Please evaluate two settings: changing the trigger position at inference time while keeping the adapter fixed, and separately training Prefix, Infix, Suffix, and Random-Position adapters. For each setting, report the ASR before and after defense, along with the cosine similarity of activation directions, grouped by trigger position. Defense effectiveness should only be compared at positions where the attack can successfully trigger before defense. Could the authors also compare the estimated directions and causal intervention results using the Last Prompt Token, a shared end anchor, and Token Pooling? This would help rule out the effects of measurement position.

  3. The authors have already reported defense ASR on Mistral-7B. Could they also repeat the cross-trigger activation direction matrix and direction ablation experiments from Finding 1 on at least Mistral and another non-Llama model, rather than only reporting the final defense performance?

  4. Please provide the original input pair indices for Figure 2(b), the model weights or cached activations, and a calculation script that exactly matches Equation (17) in Appendix D. The paper uses 60 input pairs per adapter and calculates cosine similarity after concatenating the raw mean activation shifts from layers 24–30. However, the public code in surrogate_test.py uses only 12 pairs and averages the absolute cosine similarities across layers 16–30 after normalizing each layer to unit length. These two methods do not measure the same statistic and should not be treated as equivalent. Please also clarify whether overlap between the training data of real victims and surrogates is considered allowed prior knowledge. In addition, please confirm whether the prompts used for direction estimation, recovery training, and final testing are independent of each other.

Reviewer_VgHhOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates backdoor attacks in large language models through the lens of two distinct mechanistic pathways: insertion and diversion. The authors propose that backdoors do not operate as monolithic triggers but instead exploit either the injection of malicious content into the generation stream or the redirection of model attention away from safety-relevant tokens. They develop a diagnostic framework to distinguish between these mechanisms using probing classifiers and activation patching across multiple model families. Experiments reveal that different backdoor training recipes preferentially induce one mechanism over the other, with significant implications for detection and defense. The paper further demonstrates that defenses effective against insertion-based backdoors often fail against diversion-based ones, highlighting the need for mechanism-aware security evaluations.

Critical strengths

The paper’s primary strength is the conceptual decomposition of backdoor behavior into insertion and diversion pathways, which provides a more granular taxonomy than treating backdoors as monolithic threats. This reframing directly enables mechanism-aware security evaluation rather than binary detection. Building on this, the diagnostic framework combining probing classifiers with activation patching offers causal validation of the proposed mechanisms across diverse training setups, strengthening confidence in their generalizability beyond specific attack recipes.

Finally, the empirical demonstration that standard defenses succeed against one pathway but fail against the other concretely shows that aggregate success rates are inadequate for security assessment, urging adoption of mechanism-stratified benchmarks.

Critical weaknesses

The primary weakness is the limited scope of backdoor training methods used to instantiate the two mechanisms. The paper relies on a narrow set of poisoning strategies, leaving open whether insertion and diversion exhaustively capture all possible backdoor behaviors or merely represent two points in a broader spectrum. Therefore, the generality of the proposed taxonomy remains unproven, and future work may uncover additional pathways that challenge the current dichotomy.

In addition to taxonomic completeness, the second weakness concerns the scalability of the diagnostic framework to larger models and longer contexts. The probing and patching experiments are conducted on relatively small-scale setups, raising questions about whether the observed mechanistic distinctions persist as model capacity and sequence length increase. Without validation at scale, practitioners cannot confidently apply these diagnostics to production-grade systems where computational constraints and architectural differences may alter backdoor manifestations.

Compounding these limitations, the third weakness is the absence of proactive defense proposals tailored to each mechanism. While the paper effectively exposes defensive gaps, it stops short of designing or evaluating countermeasures specifically targeting insertion versus diversion. Consequently, the work identifies a problem space without providing actionable solutions, limiting its immediate utility for defenders who must secure models today rather than await future research.

Questions for authors
  1. Have you tested whether the insertion/diversion dichotomy holds under alternative backdoor training paradigms such as instruction tuning poisoning, RLHF manipulation, or data-free weight perturbation? Expanding the scope of instantiation would strengthen claims about the taxonomy’s completeness.

  2. Furthermore, does the diagnostic framework remain effective when applied to models with significantly larger parameter counts or extended context windows, and have you observed any qualitative shifts in mechanism prevalence as scale increases? Such evidence would clarify the operational boundaries of your approach.

  3. Finally, can you sketch or preliminarily evaluate even simple defense heuristics that are explicitly conditioned on the detected mechanism, such as attention masking for diversion or output filtering for insertion? Even negative results here would help define the solution space and guide subsequent work.

Reviewer_wyFdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates why existing defenses vary in their effectiveness against LLM backdoors, and whether understanding the underlying mechanisms enables repair without knowing the attacker’s trigger.

The authors distinguish backdoors that insert abnormal behavioral directions from those that divert existing model circuits. They propose TRIAGE, which uses invented-trigger reference backdoors to diagnose the mechanism and select a corresponding repair.

Experiments show reduced attack success across three backdoor objectives. However, capability loss, reliance on continued activation interventions, and reduced effectiveness against adaptive attacks remain limitations.

Critical strengths
  1. The paper connects the insert/divert distinction to different repair strategies, offering an explanation for why suppressing an abnormal direction may work for some backdoors but interfere with legitimate behavior for others.

  2. TRIAGE uses invented-trigger references to mitigate backdoors without access to the suspect’s actual trigger, and extensive experiments across varied attack settings and model architectures support the effectiveness of this approach.

Critical weaknesses
  1. The paper does not fully explain how independent causal labels were established for all 26 adapters. It also lacks comprehensive comparisons of alternative repair strategies at matched utility, making the specific contribution of mechanism-based routing difficult to isolate.

  2. The paper does not sufficiently explain how repair thresholds and attenuation strengths are selected without feedback from evaluations using the actual trigger. This leaves uncertainty about how the method should be configured for previously unseen adapters under the stated trigger-free setting.

Questions for authors
  1. Independent validation of mechanism labels.

    The evaluated insert/divert labels closely track attack objectives, leaving uncertainty about whether the router identifies underlying mechanisms or task-specific differences. What independent causal evidence supports these labels and distinguishes mechanism identification from differences in attack objectives or training procedures?

  2. Contribution of mechanism-matched repair.

    Although component analyses are provided, the comparisons do not fully isolate the benefit of selecting a repair strategy based on the diagnosed mechanism. Could controlled comparisons with a single repair strategy applied to all adapters, under matched utility and training budgets, clarify the contribution of mechanism-based routing?

  3. Parameter selection without real triggers.

    Repair thresholds and attenuation strengths affect defense effectiveness and retained utility, yet their selection without actual-trigger feedback is not fully explained. What reproducible procedure would allow these parameters to be configured for a previously unseen adapter under the stated trigger-free assumptions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.