acceptodds
Under review as a conference paper at ICLR 2027

The 48 Laws of LLM Power: Benchmarking and Mitigating Machiavellian Agent Advice

Abstract

Large language models (LLMs) increasingly act as strategic advisers in social and professional settings, creating a safety risk poorly captured by benchmarks focused on direct model-to-user manipulation. We introduce 48LawsBench, which operationalizes Robert Greene’s The 48 Laws of Power – a distillation of three millennia of power seeking – to systematically evaluate whether LLMs act as harmful Machiavellian advisers. We test these laws at different levels of explicitness – from direct requests for unethical tactics to ordinary goals where such tactics are only implicit – distinguishing compliance from cases where models themselves infer harmful strategies. We evaluate 8 commercial and 8 open-weight LLMs and uncover striking differences: without safeguards, the share of responses using high-risk tactics ranges from ∼10% to ∼67% across models. For example, Gemini often recommends concealment, credit appropriation, dependency, and adversarial neutralization, whereas Grok does so substantially less often. Among several recurring cross-model failures, Law 15, “Crush Your Enemy Totally,” is especially persistent despite being judged highly unethical in isolation. Simple prompt safeguards reduce such behavior, but sometimes at the expense of advice utility. We thus introduce GreeneGuard, an infer–audit–rewrite safeguard that jointly targets safety and utility. It further reduces harmful strategic advice while preserving or improving usefulness, addressing residual failures left by simple prompting. Our results show that reducing harmful strategic advice need not come at the expense of effective goal-directed assistance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
social safety, manipulation, prompt safeguards

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CQ2nOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents 48LawsBench, which assesses whether large language models would recommend means to manipulate or harm third parties when providing strategic advice based on the "48 Rules of Power", and distinguishes between three scenarios: explicit requests, implicit confrontations, and neutral targets. Overall, the topic selection of the paper is of practical significance, the reasoning is clear, the writing is easy to understand, and the experimental coverage is relatively wide. However, the effectiveness of the evaluation indicators and the generalization ability of the mitigation methods still need more thorough verification.

Critical strengths
  1. The topic is of practical significance and offers valuable insights for the security assessment of large language models.
  2. The paper demonstrates sound overall logic and a clear structure, with the research motivation and key contributions explicitly stated.
  3. The experiments cover a variety of models, and the analysis of the results is comprehensive.
Critical weaknesses
  1. The evaluation relies heavily on automated scoring, and the reliability of the results requires further verification.
  2. Comparisons with other methods and ablation studies are insufficient, and the sources of the key improvements remain unclear.
  3. The experimental scenarios are somewhat limited, and the method's effectiveness in broader applications remains to be verified.
Questions for authors
  1. Provide results regarding the consistency between human evaluation and automated scoring, and clarify whether the main conclusions remain robust when the scoring model is changed?
  2. Include a comparison of methods under the same inference budget, as well as ablation studies on key modules, to pinpoint the source of the performance improvements?
  3. Does the method remain effective across a wider range of models, diverse scenarios, and situations where target rules are not pre-specified?
4 comments

Thank you for your detailed feedback! We have been working intensively on strengthening the paper along all three directions you identified:

  • W1/Q1 (Evaluation reliability): We expand human validation of LES/GUS scoring, rescore all evaluated models using a common judge, and conduct cross-judge calibration on identical responses. We also add independent PSRL validation with two human raters and a separate two-reviewer appraisal of advice actionability, willingness to follow, and expected effectiveness.

  • W2/Q2 (Comparisons and ablations): We add comparisons against an alternative iterative-refinement baseline, and evaluation with fixed-length model advice (to control for differences in advice-generation length and its associated output-token costs). We further investigate reasoning effort, model size and family, provide more detailed results for differences among prompt safeguards, and we analyze execution traces in more detail to clarify where improvements originate.

  • W3/Q3 (Generalization): We extend GreeneGuard with automatic identification of relevant high-risk laws, removing the requirement that target laws be specified in advance. We additionally evaluate the resulting system on mixed user traffic, including requests entirely unrelated to strategic advice, legitimate advice, and high-risk advice.

Could you please confirm whether these additions constitute the comparisons, validation, and generalization tests you had in mind? In particular, would you require any additional specific baseline or component-level ablation beyond those described above?

If we include these additions in the revision, would you consider revising your recommendation toward acceptance? Your feedback on any remaining methodological requirement would be extremely valuable before we finalize the revision.

I compared the current version with the original submission and found no changes in the main text or appendices. I also could not find the results of the additional experiments mentioned in your response. It remains unclear how closely the automated LES and GUS agree with human ratings, and whether GreeneGuard outperforms the iterative-refinement baseline under the same inference budget. Given how directly these questions affect the paper’s main conclusions, I am lowering my rating from 2 to 1.

We would like to clarify a misunderstanding: we have not yet submitted our revised manuscript or the results of the additional experiments.

Our previous message was intended as a preliminary update on the work underway, and to solicit your feedback before finalizing the revision.

The scope of the additional experiments and analyses has been substantial, which is why we have taken additional time to complete and carefully validate them. We are now finalizing the revised manuscript and rebuttal, and will submit both shortly.

We would be very grateful if you could reconsider your assessment after reviewing the forthcoming revision, as the currently available manuscript does not yet reflect any of these additions.

We thank you for your constructive feedback. We have revised the manuscript extensively, addressing all the comments through additional experiments, analyses, and clarifications.

New or substantially revised content is highlighted in yellow, with margin tags indicating the corresponding reviewer and comment (Q = Question, W = Weakness). To accommodate the additions, we slightly condensed some existing text.

W1. Reliability of automated evaluation

We introduce five human-validation stages (Section 2.4, Figure 2, Appendix D), alongside extensive judge-robustness analyses (Section 4.4, Appendix F.3, Figures 13–23).

Q1. Human versus automated scoring

We report human–automated scoring agreement in Appendix D.3, with additional human validation results in Appendices D.1, D.2, and D.4. The corresponding questionnaires are provided in Appendix D.6.

Q1. Robustness to changing the judge

We address this in Section 4.4 (Figure 4) and Appendix F.3 (Figures 13–23). The tested cross-model ordering persists across judges, despite differences in their absolute scores.

W2. Comparison with other methods

We add a comparison with Self-Refine in Appendix F.4.6 (Figure 28), evaluated under two judge settings. GreeneGuard more effectively suppresses high-LES responses for the most problematic laws while preserving high GUS.

W2/Q2. Ablation studies.

We substantially extend the ablations in Appendix F.4 (Figures 24–28), including analyses of judge reasoning effort, model size, model generation, output length, and baseline instructions.

We further discuss the underlying patterns in Section 4.5 (p. 9) and Appendix F.5 (Figure 29).

Q2. Equal inference budgets

Appendix F.4.2 (Figure 26) adds a response-length sensitivity analysis where we force the answer length to be between 150-200 tokens.

We also analyze inference costs in Appendix F.4.3 (Tables 6–7) and in Appendix F.4.6 (Figure 28) where we investigate GreeneGuard and Self-Refine.

W3/Q3. Broader scenarios and unspecified target laws

We broaden the evaluation across models and settings (Appendices F.3–F.5, Figures 13–29). More importantly, Appendix F.6 (Figures 30–31) introduces a multi-law GreeneGuard extension that identifies and mitigates active high-risk laws without specifying them in advance for each scenario. Across 20 scenarios combining multiple laws, 13 responses satisfy the joint risk and utility criterion.

Reviewer_dB1jOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies LLM safety under a threat model the authors call the Machiavellian adviser: the model does not manipulate the user directly, and is not pursuing its own objective, but helps the user acquire, exercise, or preserve social power over a third party. This is positioned as orthogonal to MACHIAVELLI (agents harming others while pursuing their own goals) and to persuasion/deception benchmarks (the model manipulating its interlocutor).

Critical strengths
  1. The threat model is a real gap, and the paper makes it checkable rather than rhetorical. Advisor-mediated power seeking (the model helps A dominate B) is distinct from both agent-goal-directed harm and direct user manipulation. Table 1 compares eight prior benchmarks along six explicit axes (advisor-mediated, implicit emergence, strategy-specific, power acquisition, mitigation, goal preservation), turning a positioning claim into a verifiable table. This is the contribution most worth keeping.

  2. Using Greene as an external taxonomy avoids the circularity of self-authored taxonomies. The strategy space is drawn from a historically and culturally broad compendium spanning Machiavelli, Clausewitz, and Sun Tzu, and covers dependency creation, scarcity, choice control, status management, coalition disruption, intimidation, and rival elimination — considerably wider than deception- or dark-pattern-centric taxonomies. Each law is rewritten into a decidable operational behavior (e.g., Law 15 = "neutralize an adversary so completely that recovery is impossible"), which is more rigorous than most "borrow a classic framework" papers.

Critical weaknesses
  1. The judge pool is uncalibrated, and the judge assignment is logically incompatible with the paper's headline claims. Three judges from three providers are used: GPT-5 and GPT-5.4 are judged by Gemini-3.1-pro-preview; GPT-5.4-mini by Grok-4.6; everything else, including all Gemini models, by GPT-5.5. The paper reports no cross-judge calibration experiment of any kind — no shared scored subset, no inter-judge correlation, no systematic-offset analysis. More seriously, the design is internally inconsistent with the claims. The authors write that this assignment provides "a constant evaluator within each model's comparisons" (App. D.1), which is an admission that the design licenses within-model before/after comparisons only. Yet the abstract, Figure 1, and §4.1 are all built on cross-model comparisons ("Gemini often recommends … whereas Grok does so substantially less often"). The situation is aggravated by a circularity: OpenAI outputs are judged by a Gemini model while Gemini outputs are judged by a GPT model, and the paper's most emphasized ranking is precisely between these two families. I am not suggesting anything improper — but this is a confound that must be disclosed and stress-tested, and it currently is neither.

  2. Output length is a strong, uncontrolled confound. LES asks whether advice realizes a tactic; longer, more concrete, more step-structured advice simply has more opportunity to contain any given tactical element. Table 5 shows generation lengths that differ by an order of magnitude: Qwen3.5-9B median completion 2,310.5 tokens, Grok 1,090.5, Gemma 645, Llama-3.1-8B 233. The authors cite Dubois et al. (2024) on length bias in B.2.5 and then do not control for it anywhere.

  3. The baseline advice instruction itself induces the failures being measured. The instruction in Appendix B.2.1 is: "Your response should be practical, strategic, and should maximize effectiveness while being as concise as possible." This directly instructs the model to select the most effective means available, which inflates the unmitigated baseline and amplifies between-model spread. The "unmitigated" condition is therefore an effectiveness-maximizing adversarial condition, not default deployment behavior — yet the abstract presents its numbers without this caveat ("without safeguards, the share of responses using high-risk tactics ranges from ~10% to ~67%").

  4. The PSRL assignment is unvalidated but determines the entire dependent variable. The five-level risk assignment is author-assigned with no inter-annotator agreement and no sensitivity analysis, yet "high-risk = PSRL 3–4" is the definition on which every reported number depends. Several boundary assignments are contestable: Law 11 (create dependency) = 4 while Law 16 (use scarcity to increase perceived value) = 2; Law 4 (restrict what one reveals) = 2 while Law 3 (hide one's real objective) = 3. Moving these boundaries changes the high-risk law set and hence all results.

Questions for authors

See the weakness

2 comments

Thank you for your detailed review! We have been working intensively on addressing each of your four weaknesses through additional experiments and methodological clarifications:

  • W1 (Judge confounding): We rescore all evaluated models using a single common judge. Independently, multiple judges score identical responses, allowing us to quantify agreement, scoring offsets, and overall robustness of cross-model rankings.
  • W2 (Output length): We add length-adjusted analyses and controlled evaluations with matched response lengths, testing whether model-level differences persist after accounting for this factor.
  • W3 (Baseline instructions): We compare our original effectiveness-oriented prompt against a neutral advice instruction on identical scenarios. We also clarify that the original prompt requests effectiveness, not manipulative tactics, and explicitly qualify the interpretation of baseline rates.
  • W4 (PSRL assignments): We explain the classification rationale and validate all 48 laws using two independent human raters (Krippendorff's α = 0.68; Kendall's W = 0.83).

Our key question is whether this set of additions covers the methodological requirements underlying all four weaknesses. Are there any specific missing controls or experiments that you would still consider necessary?

More importantly, if these methodological concerns are addressed through the above analyses, would you consider revising your recommendation toward acceptance? We would greatly appreciate your guidance before finalizing the revision.

We thank you for your constructive feedback. We have revised the manuscript extensively, addressing all the comments through additional experiments, analyses, and clarifications.

New or substantially revised content is highlighted in yellow, with margin tags indicating the corresponding reviewer and comment (Q = Question, W = Weakness). To accommodate the additions, we slightly condensed some existing text.

1. Judge calibration and cross-model comparisons

We address this extensively in Section 4.4 (Figure 4, p. 9) and Appendix F.3 (Figures 13–23), including cross-judge agreement at the law and response levels, systematic scoring offsets, and additional fixed-judge comparisons. The tested model ordering remains consistent across judges, although absolute scores and the magnitude of differences vary.

2. Output-length confound

We add an output-length sensitivity analysis in Appendix F.4.2 (Figure 26). Shortening GPT-5 advice changes mean LES only modestly (40.6 to 38.4), although individual scenarios may exhibit larger differences.

3. Effectiveness-oriented baseline

We introduce a vanilla baseline without explicit effectiveness emphasis (Appendix F.4.4, Figure 27). High-risk responses decrease from 19/60 to 11/60 but remain present, while GUS stays near 4. Thus, effectiveness emphasis increases high-risk behavior but does not cause it by itself.

Please note that our original unmitigated baseline merely requests effective advice, not manipulative tactics; models remain free to choose constructive or harmful means.

4. PSRL validation

We introduce independent human validation of the PSRL assignments in Appendix D.1. The raters, blinded to our original labels, achieve Krippendorff's α = 0.68 and Kendall's W = 0.83, supporting substantial agreement. We also update Figure 2 (p. 3) and summarize the validation procedure in Section 2.4 (p. 4).

Reviewer_jNGaOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces the risk of LLMs providing machiavellian advice to users enabling harmful power seeking. They operationalize a propensity evaluation drawing on the book “48 laws of power” by Robert Greene, creating a comprehensive benchmark. They evaluate a set of popular open and closed LLMs on this benchmark. They further study several safeguarding techniques that reduce the propensity to give machiavellian advice while retaining usefulness for the user’s goals.

Critical strengths
  1. The contributions seem comprehensive: the authors conceptualize a novel risk (machiavellian advice), find a clever way to operationalize it, study potential ways to mitigate it (prompting and re-writing using their Law Embodiment Score and Goal Utility Score), and evaluate popular closed- and open-source llms on this benchmark.
  2. LLMs as machiavellian advisors seems like a realistic and useful setup to evaluate.
  3. Evaluation regimes with differing levels of explicitness is a great idea. Evaluating propensity towards some misaligned behavior often depends strongly on the amount of nudging the evaluators do. Making this variable explicit helps more precisely state what a benchmark tells about the model.
  4. The fact that a law is judged to be unethical, but is still applied highlights is an interesting and surprising finding. It highlights inconsistency in LLM alignment, intuitively a discrepancy between what they say and what they do. Probably deserves more attention.
Critical weaknesses
  1. The evaluation of different llms does not seem to tell us much about the reasons behind this propensity and trends for its expression.

  2. While the contributions of this paper are already enough to justify acceptance, there is a certain discrepancy between the stated risk and what the benchmark of this sort can achieve. L044 states: “LLMs as Machiavellian advisers: strategic intermediaries that help users acquire, exercise, or preserve social power over third parties.” 48LawsBench measures the propensity of LLMs to give such advice, but it does not cover how often and when humans would accept this advice, and whether this advice would be useful in helping them gain power.

  3. Since the paper conceptually introduces machiavellian advising as a risk - understanding the threat model in detail is necessary to motivate the reader’s attention. In simpler terms: you propose to measure something about the model. Why should anybody care?

  4. Understanding the threat model also informs understanding the use of safeguards. Are the prompt-based safeguards intended to be used by the user? What if the user is actively seeking machiavellian advice? If it is intended to be used in the system prompt by the model provider - one potentially needs to measure its influence on unrelated user traffic. Same goes for GreeneGuard.

Questions for authors

Please see Weaknesses.

2 comments

Thank you for your constructive and encouraging review! We have been working intensively on addressing all four points, and have prepared the following additions to the revision:

  • Q1 (Explaining model behavior): New experiments investigating the impacts of factors such as model size, generation, and reasoning effort.

  • Q2 (Real-world relevance): A new human study with two independent reviewers, assessing willingness to follow the advice, its practical actionability, and expected effectiveness if followed. We also provide the annotation methodology and form.

  • Q3 (Threat model): An explicit three-party threat model (LLM, user, affected third party), supported by quantitative evidence from existing empirical studies on AI-advice uptake, behavioral influence, and third-party consequences.

  • Q4 (Safeguards): A provider-controlled deployment architecture with automatic risk detection and selective mitigation, together with an evaluation covering unrelated requests.

Could you please let us know whether these four additions, in principle, adequately address your respective questions? In particular, is there any specific aspect of the threat model or its validation that you believe still needs to be covered?

We would greatly appreciate your guidance while finalizing the revision!

We thank you for your constructive feedback. We have revised the manuscript extensively, addressing all the comments through additional experiments, analyses, and clarifications.

New or substantially revised content is highlighted in yellow, with margin tags indicating the corresponding reviewer and comment (Q = Question, W = Weakness). To accommodate the additions, we slightly condensed some existing text.

1. The evaluation of different llms does not seem to tell us much about the reasons behind this propensity and trends for its expression.

We address this through additional experiments and ablations (Appendix F.4, Figures 24–28), together with a dedicated discussion of the observed patterns and their possible explanations (Section 4.5, p. 9; Appendix F.5, Figure 29).

2. (...) 48LawsBench (...) does not cover how often and when humans would accept this advice, and whether this advice would be useful in helping them gain power.

We introduce Practical Relevance Validation, assessing willingness to follow advice, actionability, and expected effectiveness (Section 2.4, p. 4; Appendix D.4; questionnaire in Appendix D.6).

3. Threat model and its relevance

We clarify our three-party threat model in Section 2 (p. 2) and provide a detailed discussion with supporting evidence in Appendix A (Table 2, p. 15). This connects the benchmark to the potential real-world consequences of LLM-generated advice.

4. Safeguard deployment and unrelated traffic

We outline a potential provider-controlled deployment architecture in Appendix E.4. We implement and evaluate its initial advice-classification component on strategic and unrelated user traffic (Appendix F.4.5, Table 8), achieving 98–100% classification accuracy across seven models.

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback. We have extensively revised the manuscript, addressing their comments through additional experiments, analyses, and clarifications.

The main additions include:

  1. expanded human validation of PSRL, LES/GUS, and advice relevance, including questionnaire forms,
  2. extensive cross-judge robustness and fixed-judge comparisons,
  3. new baselines and ablations, including Self-Refine and inference costs,
  4. analyses of output length, reasoning effort, model size and generation, and baseline instructions,
  5. deeper analyses of model-specific failures, safeguard differences, and effectiveness–safety trade-offs,
  6. generalized GreeneGuard handling arbitrary combinations of laws without pre-specified targets,
  7. advice classification on mixed user traffic and a proposed safeguard deployment architecture.

New or substantially revised content is highlighted in yellow, with margin tags identifying the corresponding reviewer and comment (Q = Question, W = Weakness). To accommodate these additions, we slightly condensed some existing text.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.