acceptodds
Withdrawn from ICLR 2027

Defending Large Language Models Against Diverse Jailbreak Attacks via Multi-Concept Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Despite extensive safety alignment efforts, Large Language Models (LLMs) remain highly susceptible to jailbreak attacks. Although various defenses have been proposed, their effectiveness remains limited against emerging or adaptive threats, largely because the underlying mechanisms of jailbreak attacks remain poorly understood. To elucidate these mechanisms, recent studies leverage the Linear Representation Hypothesis to extract jailbreak concepts from LLM representations. However, our analysis reveals that no single jailbreak concept generalizes across diverse jailbreak techniques, as different attack strategies exhibit distinct representation patterns. To address this limitation, we propose Multi-Concept Detection (MCD), a lightweight defense framework that robustly identifies jailbreak attempts by extracting their multi-concept features. Extensive experiments on five open-source LLMs demonstrate that MCD consistently achieves superior detection performance and remains robust against a wide range of attack strategies.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Large Language Models, Jailbreak Defense, Representation Engineering, Concept Probing and Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_LRCWOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes multi-concept detection, a jailbreak prompt detector. The authors conduct experiments on several models, datasets, and jailbreak attacks.

Critical strengths
  1. Table 1 covers a wide range of jailbreak attack methods, and detection baselines.
Critical weaknesses
  1. The idea is not novel. Jailbreak detector has been a well-studied topic. It has also been works designing jailbreak defenses based on internal activations/features.

  2. The defense lacks practical considerations. How much additional time/compute does it require to run the defense during deployment? The largest LLM used in this paper is merely 13B. And the LLMs are generally out-of-dated. In summary, I have serious doubt about the method's performance and effiency on large-scale, latest LLMs such as Qwen3.6-35B.

  3. The most concerning shortcoming of the proposed method is the need to enumerate exisitng jailbreak attacks, and the authors did not conduct evaluation on unseen/OOD jailbreak attacks. In practise, the attackers very likely to use attack algorithms never published on papers. As a result, it is important to show that the detector is still effective on such attacks.

Questions for authors

Please see weakness.

Reviewer_V6hUOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies representation-based defenses against LLM jailbreak attacks based on the Linear Representation Hypothesis (LRH). The authors propose a Multi-Concept Detection (MCD) method, which constructs a multi-concept set comprising the harmfulness direction and multiple jailbreak directions, projects prompts onto these directions, and trains a lightweight classifier. It claims that MCD consistently achieves superior detection performance and remains robust against a wide range of attack strategies compared to baselines.

Critical strengths
  1. The motivation is clear and practical.
  2. The proposed method is intuitive and easy to understand.
  3. The experiments are comprehensive, including 5 LLMs, 6 baselines, and 9 attack methods. It also includes external validation on XSTest and WildGuardMix.
Critical weaknesses
  1. The contributions and advantages of the proposed method are not convincing. In the abstract and experimental section, it is claimed that MCD consistently achieves superior detection performance and remains robust against a wide range of attack strategies. However, under SCAV attack (Table 4), MCD detection accuracy drops to 0.75% on Vicuna-7B. In addition, as shown in Table 1, MCD is below AlignTree on PAIR for all 5 LLM models (e.g., Vicuna-13B 84.52 vs 96.71; Llama-3-8B 88.66 vs 95.73) and on IJP for all 5 LLM models. It is also at or below AlignTree on benign prompts. In Table 2, MCD also performs poorly on benign prompts.
  2. Missing Leave-One-Attack-Family-Out evaluation. In Table 1, the multi-concept basis and the LightGBM classifier are trained using 30 prompts from all 9 jailbreak methods, and then evaluated on held-out prompts from those same 9 methods. While XSTest and WildGuardMix (Table 2) test out-of-distribution generalization, the paper's core idea, combining multiple known jailbreak concepts to generalize across diverse attacks without prior knowledge, requires a leave-one-attack-family-out (or unseen attack family) study. This will help to answer "Does the remaining K-1 concept basis still capture the held-out attack, or does MCD require training samples from every known attack family to be effective?"
Questions for authors
  1. In Eq. 2), the paper states that prompt embeddings are centered using a single anchor μb\mu_b. However, Figure 2 and Algorithm 1 & 2 (Appendix B) explicitly construct and pass two anchors (μb\mu_b and μh\mu_h). This seems to be an inconsistency. Please clarify this.
  2. How does performance change with the number of training prompts per attack and with K?
  3. What are the latency and memory overhead compared with AlignTree?
  4. In Table 1, only detection accuracy is shown. How about other criteria, e.g., TPR, TFR, F1?
  5. In the introduction, it is stated "propose a Multi-Concept Detection (MCD) framework that captures complex malicious patterns without sacrificing interpretability.". But it is not explained clearly why interpretability is not sacrificed.
Reviewer_dS7cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper reveals that single jailbreak concept generalizes poorly across diverse jailbreak techniques. To address this limitation, they propose Multi-Concept Detection (MCD), a lightweight defense framework that robustly identifies jailbreak attempts by extracting their multi-concept features. Extensive experiments are performed to demonstrate the detection generalization.

Critical strengths
  1. The preliminary analysis identifies the limited cross-attack generalization of a single jailbreak concept direction, providing a clear motivation for the proposed method.
  2. MCD is lightweight and interpretable, and is evaluated across multiple LLMs, jailbreak attacks, external datasets, and classifier architectures.
Critical weaknesses
  1. The methodological novelty appears limited. The preliminary analysis provides an interesting observation that jailbreak concept directions can be attack-specific. However, the proposed MCD framework mainly addresses this issue by concatenating multiple concept projection scores and feeding them into a standard supervised classifier. Thus, the improvement may largely result from using a richer multi-dimensional feature representation and learned feature fusion, rather than from a fundamentally new representation or generalization mechanism.
  2. The task formulation is not well aligned with the paper’s motivation. Section 4.1 focuses on generalizing jailbreak detection across diverse attacks, which is a two-class classification problem, while the method is formulated as a three-class problem over benign, harmful, and jailbreak prompts. Section 4.2.4 further merges harmful and jailbreak prompts into a single malicious category at inference, yet the experiments still report them separately. The motivation, training objective, inference rule, and evaluation protocol are therefore not fully consistent.
  3. The necessity of three-class training is insufficiently justified. If the final task is jailbreak detection or benign-vs-malicious detection, it is unclear why harmful and jailbreak prompts need to be explicitly separated during classifier training. An ablation comparing three-class and binary training is needed to demonstrate the benefit of this design.
  4. The experiments do not directly validate cross-jailbreak generalization. MCD constructs concepts and trains the classifier using samples from all nine jailbreak methods, and then evaluates on remaining samples from the same attack families. This mainly tests unseen samples of seen attacks rather than unseen jailbreak strategies. A leave-one-attack-out or train-on-subset/test-on-unseen-attacks setting would better support the motivation. The external datasets in Section 5.2.2 do not fully resolve this issue, since XSTest contains no jailbreak samples, and the relationship between WildGuardMix attacks and the training attacks is not analyzed.
  5. The contribution and necessity of the supervised classifier are unclear. Since the downstream classifier explicitly learns the relative importance and combination of multiple jailbreak-related concept scores, the current experiments cannot disentangle whether the performance improvement mainly comes from the proposed multi-concept representation or from the supervised fusion learned by the classifier. Comparisons with simple fusion strategies such as mean/max aggregation or threshold-based rules are needed. Moreover, visualizing the multi-concept feature space, analogous to Figure 1, would help verify whether it actually improves cross-attack separability.
Questions for authors

Could you specify the training protocol used for each trainable baseline?

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

Related papers

Discussion (0)

Sign in to comment.