acceptodds
Under review as a conference paper at ICLR 2027

Jailbreaks Live in Persona Space

Abstract

Understanding why aligned language models fail is central to AI safety. Jailbreaks provide a concrete instance, where adversarial prompts can shift models from refusal to harmful compliance. We study jailbreaking through the emerging view of LLMs as persona simulators, asking whether jailbreak behavior is organized within persona space. On Llama-3.1-8B and Gemma-4-31B, we find that jailbreaks are strongly structured by persona space. Individual traits predict jailbreak outcome, persona structure emerges in principal components learned solely from jailbreak activations, and persona-space projections retain most of the predictive signal in refusal and jailbreak directions. Building on this structure, we construct the PersonaJB Direction, a single direction constrained entirely to persona space that achieves the strongest average generalization to unseen attacks on both models with minimal inference overhead. Finally, steering along PersonaJB and individual trait directions induces jailbreaks at rates comparable to specialized safety directions and standard attacks. Our results establish persona space as an interpretable, predictive, and causal representation of jailbreak behavior.

Primary area:
interpretability and explainable AI
Keywords:
Jailbreaks, persona space, interpretability, activation steering, monitoring, alignment

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_EnH1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether persona-trait vectors can explain and predict jailbreak outcomes. It constructs a trait subspace and trains a logistic-regression detector, PersonaJB, on hidden-state projections onto these traits. Experiments on Llama-3.1-8B and Gemma-4-31B show improved average detection AUC on five unseen attack families. Steering and ablation experiments on Llama further show that modifying these directions can change refusal behavior.

Critical strengths
  1. The detection gains appear on both evaluated models. PersonaJB achieves higher average OOD AUC than raw-activation and dimension-matched PCA probes. This supports the usefulness of the trait representation beyond reducing the number of features.

  2. The paper tests interventions as well as predictions. Steering along PersonaJB and individual trait directions changes refusal behavior, while random-direction controls are substantially weaker. These experiments show that the identified directions can affect model behavior.

Critical weaknesses
  1. The layer-selection procedure leaves the OOD comparison difficult to assess. Appendix A.3 states that OOD attacks do not enter model selection, but Appendix C motivates Llama layer 16 using its highest average transfer AUC. It is unclear whether this was a post-hoc observation or the criterion used to select the layer. Gemma’s layer-30 choice is also not explained through a comparable analysis. More generally, the paper needs a clear procedure for choosing an extraction layer and validating trait vectors on a new model without consulting the final test attacks.

  2. The analysis does not establish whether trait scores predict outcomes beyond differences in request content. Different harmful-content categories may both produce different trait scores and have different jailbreak success rates. The paper compares attack families, but does not report category-controlled or within-request analyses. Behavior-disjoint splits do not rule out this explanation, and signal emergence at assistant-header tokens does not establish independence from content. This limits the interpretation of the results as evidence for a general persona-level account of jailbreaks. The exclusive use of traits, without evaluating role-based representations, further limits that scope.

Questions for authors
  1. Was Llama layer 16 selected before examining OOD performance, or because it achieved the highest OOD AUC? How was Gemma layer 30 selected, and were baseline layers chosen using the same procedure?

  2. Please provide, or identify in the released code, the 200 extraction questions, positive/negative system prompts, and scoring rubrics. How are response-token activations aggregated into the representation used to construct each trait vector?

  3. Please specify the train/validation/test sizes, the WildJailbreak subset used for Gemma, and the tuning ranges and selected configurations for PersonaJB, the raw-activation probe, and the PCA probe. Were standardization and PCA fitted only on the training portion of each split?

  4. Additional evidence needed for the current claim: For attack variants targeting the same original harmful request, do trait scores still distinguish successful from failed attacks? An analysis restricted to requests with both outcomes would help determine whether the detector captures attack-induced changes rather than mainly distinguishing easier requests from harder ones.

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

By "persona space" the paper means a low-dimension subspace of an LM's residual stream, defined by, roughly, 240 so-called "personality traits" directions (e.g., 'absolutist', 'progressive', 'romantic', 'stoic', ...). The paper asks whether inspecting the persona space suffices to predict or cause jailbreaks. To this aim, the authors prompt Llama3.1-8b or Gemma4-31b with jailbreak prompts (extracted from HarmBench and wrapped around jailbreak templates, or WildJailbreak for Gemma), read their residual streams at the last token before generation (at layer 16 for Llama, layer 30 for Gemma), and evaluate whether the LMs have been jailbroken using GPT-4.1-mini as an LLM-as-judge. They claim that:

  • i. Fig. 2: Projecting the residual stream onto every trait direction and turning this score into an AUC (aimed to measure if the LM will be jailbroken) results in up to > 0.85 AUC for the two tested models and the best traits, with average (over traits) AUC > 0.65.
  • ii. Fig. 3: In Llama3.1-8b, method (i) with a different metric ("succeeded failed normalized distance") separates successful jailbreak prompts from un-successful ones mostly if the activations are those of the assistant turn.
  • iii. Fig. 5: The "refusal direction" and the "jailbreak direction" (computed as the average difference of harmful (resp. jailbreaking) prompts minus harmless (resp. non-jailbreaking) prompts) have ~26 to 48% of their squared norm in the persona space, and the projections of these vectors onto the persona space barely changes their AUC at predicting jailbreaks.
  • iv. Fig. 6: A direction the authors call "PersonaJB" (learned via logistic regression over trait directions) has the best average AUC over fine unseen jailbreaks, against 13 baselines.
  • v. Fig. 8: Steering Llama along PersonaJB or individual trait directions results in jailbreak rates higher than or comparable to safety steering or prompt-based attacks.
Critical strengths

i. The hypothesis underlying the paper is interesting, and seems both plausible and important to study. More generally, finding low-dimensional subspaces of language models' internals that (approximately) suffice to predict their behavior (beyond of predicting jailbreaks) is something I find worth doing.

ii. The paper explores the phenomenon under multiple different lenses.

  • E.g., (a) whether the (low-dimensional) persona component of jailbreak prompts suffices to predict jailbreaks, (b) at which token positions the prediction signal appears, (c) whether the norm of refusal and jailbreak directions "lives" in persona space, (d) whether the persona-component of these vectors is predictive of jailbreaks, and (e) whether steering along persona directions increases jailbreaks.
  • Some of these lenses come with proper controls (e.g., comparing persona steering with steering in random directions in the full space and in the persona space) and numerous baselines (e.g., 13 jailbreak detectors baselines, cf. Fig. 6).
Critical weaknesses

i. The prose of the paper is often heavily underspecified. Consequently, results or claims are hard to interpret or verify. Some examples of missing details are:

  • On which data the sign of each trait direction is chosen (e.g., 'avoidant' is supposed to predict refusal, yet it is reported with AUC 0.8, not 0.2; cf. Tables 1 and 5);
  • The metric in Fig 3, only (under)specified in words;
  • The layers and token positions of some of the 13 baselines in Fig. 6 (thus, not allowing a full comparison with the paper's methods);
  • How the five ood jailbreak attacks are produced; Disclaimer: it is possible that some of these details can be inferred from the code or are present in the Appendix. From a few checks of the Appendix they are not, nor are they mentioned or referred to in the main text.

ii. Some of results and their interpretation could be inflated (in part, due to the absence of baselines). For example:

  • In the detection experiments (Fig. 2), the authors do not test what is the AUC of random or non-persona directions. Relatedly, they do not measure the jailbreak detectors on non-jailbreak prompts.
  • Steering along the PersonaJB results in worse ASR in both ood benchmarks, relative to the refusal (in JBB) and jailbreak (in WildJB) baseline (see Fig. 8).
  • For Llama, the authors choose layer 16 because it generalizes better at detecting jailbreaks, but they tested generalization only on the same five unseen attacks their selection of layer is based on. Moreover, some baselines are also tested at layer 16, and not relative to their best layer (for other baselines the layer is not reported).
  • Llama and Gemma are evaluated on different data. In particular, Llama uses HarmBench only, whereas Gemma uses a mix of HarmBench and WildJailbreak (because, according to the authors, HarmBench prompts rarely jailbreak Gemma). This might imply that the comparison in Fig. 2 confounds a difference in models with a difference in data. As a corroboration of this, notice that on HarmBench alone Gemma's "forgiving" direction is the best one, while being the worst on the mixture.
  • Some of the most important results (e.g., Fig. 3, Fig. 8) hold only for one model, Llama3.1-8b. The paper does not report the results for the only other model tested, Gemma4-31b, raising the question of whether they are negative or absent. Despite this, the authors state their findings in general. (Relatedly, the validation of the layers is also done for Llama only, differently from what the main text implies.)
  • To corroborate that the findings are specific about the persona space, they should be tested against another (reasonable) ~240-dimensional subspace.
Questions for authors

My questions are implicitly stated in my assessment of the critical weaknesses. For example:

  • Can you specify the missing details mentioned in weakness (i)?
  • Why are Gemma results missing?
  • How should we compare the two models you tested, in light of weakness (ii.4)?

Future work:

  • What would happen under the additional, suggested controls?
  • What would happen in other models of similar size?
Reviewer_vJFhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This work analyzes jailbreaks through the persona space, which is defined as the space of human-interpretable persona traits. The analysis addresses both individual persona traits, as well as the persona subspace as a whole. Through a series of experiments on two models, the authors attempt to establish that the persona space is an interpretable, predictive, and causal representation of jailbreak behavior. Interpretability is found by seeing if principal components of jailbreaks lie along persona directions. Predictivity is measured by projecting model representations onto the persona space and training classifiers. Causality is assessed through steering experiments.

Critical strengths

(S1) Well written. It is easy to trace the paper's claims, and how those claims come about from the results & experiments.

(S2) Their experimental finding that persona space can predict jailbreaks. Their experiments show that the persona space (either the entire space, or individual traits) can be used to effectively predict jailbreak success, and detection based on the persona space has the best OOD generalization.

Critical weaknesses

Overstated Claims. The focus of this paper is on analysis of the jailbreaking problem. There's no new method/dataset introduced (which by itself, is fine), so the paper should be judged on the insights presented.

The paper intends to establish the persona space as an interpretable, predictive, and causal representation of jailbreak behavior. The first two aspects are addressed in the paper (though the interpretability of the persona space was shown in prior work). However, the "causal" part is not convincing. Some reasons for this are:

  1. The paragraph starting at line 247 tries to show that the persona space captures harmfulness and jailbreak success. However, only between 30-50% of the squared norm of the vectors for harmfulness and jailbreak-success is captured by the persona space. This leaves a significant component that is not captured, which the paper seems to ignore. Should 30-50% be considered as being enough to infer that "jailbreaks live in persona space"?

  2. The causality experiments (Section 5) are done by steering a single model, on a single layer. Is this enough to establish causality? Aside from steering, I would expect the paper to try establishing that persona representations are the mechanism responsible for naturally occurring jailbreaks.

Given the lack of methodological novelty, I am evaluating the paper based on the insights it offers. I do not think causality is compellingly established, so the remaining insights are around predictability (since I think interpretability mainly comes from the prior work which introduced the persona space). I don't think the predictability insights (while nice) contribute sufficient new knowledge. I encourage the authors to consider: (1) conducting new/additional experiments to establish causality, and/or (2) arguing why my conclusions re causality are not correct, and/or (3) re-framing paper away from causality and focus more on predictability.

Questions for authors
  1. In Section 5, is it reasonable to steer over multiple steering strengths and declare success if any of them work? Or is it more reasonable to use a validation set to determine the steering strength hyperparameter and then report test results?

  2. I think Section 3.3's finding that "persona space preserves jailbreak-predictive information" is not so surprising, especially given the results presented in Section 3.1. What happens if you remove the persona component? If jailbreak prediction degrades significantly, this may establish that the persona space is necessary for predicting/decomposing jailbreaks.

  3. The results in the paper are presented on two models. Have the authors checked if their findings on other models (e.g., a Qwen model)? This would certainly help strengthen the generalizability of the claims.

Reviewer_4VuKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores the role of "persona space" in LLM jailbreaking, showing that jailbreak and refusal signals substantially concentrate on persona axes and proposing the PersonaJB Direction as a single-dot-product pre-generation linear detector, which demonstrates competitive out-of-distribution detection AUC across unseen attack families.

Critical strengths
  1. This paper introduces human-interpretable persona space into LLM safety and links jailbreak vulnerabilities to pre-generation persona representations, which bridges offensive representation steering and defensive monitoring within an auditable, unified geometric framework.
  2. The proposed PersonaJB Direction detector is constrained entirely to persona space and executes via a single dot product on internal activations, which delivers strong out-of-distribution detection effectiveness while maintaining negligible inference overhead.
Critical weaknesses
  1. The random baseline controls in Section 4 only reflect the average performance of randomly sampled directions, which fails to rule out whether a simple random search under an equal budget could find directions with comparable predictive power. In representation-level probing, linear performance is inherently governed by extreme-value statistics, yet the paper lacks analysis on the top-k or maximum performance distribution over random vectors. Could authors report the empirical performance distribution across random directions and provide equal-budget search curves to demonstrate that the PersonaJB Direction significantly outperforms the best candidates found by generic random search?

  2. The evaluation does not sufficiently isolate the specific structural contribution of the 240 persona traits from generic supervised dimensionality reduction. Moreover, under strictly matched subspace dimension, training data, and regularization, it remains unclear whether an arbitrary random orthonormal basis, a PCA-derived subspace, or a randomly rotated trait space could achieve comparable OOD detection AUC. Could author incorporate dimension-matched random-subspace baselines to substantiate that semantic persona directions provide unique geometric inductive bias beyond standard low-rank linear probing?

  3. In Section 5, the causal steering evaluation relies exclusively on a binary compliance metric without generation-quality gating or side-effect profiling on benign inputs. Representation steering along unconstrained internal directions frequently induces severe side effects, which can easily trick simple rule-based or automated judges into logging false-positive jailbreaks.

Questions for authors
  1. Could authors report the empirical performance distribution across random directions and provide equal-budget search curves to demonstrate that the PersonaJB Direction significantly outperforms the best candidates found by generic random search?

  2. Could author incorporate dimension-matched random-subspace baselines to substantiate that semantic persona directions provide unique geometric inductive bias beyond standard low-rank linear probing?

  3. Could authors give out generation quality metrics (such as perplexity or coherence scoring), evaluate benign false refusal rates under steering, and present trade-off curves across steering coefficients to confirm that the observed ASR gains represent genuine, high-quality jailbreaks rather than pseudo-compliance caused by text degradation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.