acceptodds
Under review as a conference paper at ICLR 2027

KBF: Knowledge Boundary as Fingerprint for Language Model and Black-Box API Auditing

Abstract

Relay and reseller APIs mediate access to large language models (LLMs), but users cannot directly verify which model serves them. We introduce KBF, a black-box auditing protocol based on stable factual recall near the knowledge boundary, including repeatable wrong answers. KBF generates benign, renewable probes and calibrates audit decisions against reference self-variation. Across 16 production endpoints, KBF detects all 155 economically relevant substitutions without rejecting any of the 16 same-reference controls. KBF remains robust to deployment variation and reaches 95% TPR at a substitution rate as low as 15% in mixed-routing simulations. Field audits flag 7 of 28 endpoints across six platforms as statistically inconsistent with their references. After reference enrollment, even GPT-6 Astra costs only approximately $0.67 per online audit at the recorded API prices.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
black-box LLM auditing, model fingerprinting, knowledge boundary, model substitution, API relay

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_EygSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Knowledge-Boundary Fingerprinting (KBF), a black-box auditing framework for detecting potential model substitution in third-party LLM APIs. The key observation is that factual recall near a model's knowledge boundary can provide a distinctive behavioral fingerprint, including stable but incorrect answers. The proposed framework constructs reference-specific probes, filters them based on response consistency and optional contrastive comparisons, and applies a calibrated binomial test to determine whether a suspect endpoint is statistically consistent with the claimed reference model.

The authors evaluate KBF across 16 production LLM endpoints, reporting successful detection of all 155 economically relevant model substitutions without false positives in 16 same-reference controls. Additional experiments examine robustness under deployment variations, partial-routing scenarios, and real-world relay APIs.

Critical strengths
  1. Practical and well-motivated problem. The paper addresses an important issue in the increasingly complex LLM API ecosystem, where users may have limited visibility into the actual models serving their requests. The black-box setting is realistic, and the proposed approach does not require access to model weights, logits, or provider-side information.

  2. Interesting use of stable factual recall. The observation that repeatable factual errors can serve as model-specific fingerprints is interesting. The distinction between common knowledge, stable boundary recall, and unstable beyond-boundary recall provides an intuitive motivation for constructing discriminative probes.

  3. Extensive empirical evaluation. The experiments cover multiple model families, price tiers, and deployment configurations. The reported detection accuracy is strong, particularly under variations in system prompts, temperature, and retrieval-augmented settings. The partial-routing experiments and real-world API audits further demonstrate the practical relevance of the approach.

Critical weaknesses
  1. Problem motivation and practical necessity. While the problem is practically relevant, I feel that the motivation could be further strengthened from a user's perspective. For instance, users may evaluate model quality through direct experience or reduce substitution risks by using official APIs. It would be helpful if the authors could further clarify the scenarios where dedicated auditing is particularly valuable.

  2. Novelty and methodological contribution. The framework builds upon several existing techniques, including factual probing, consistency filtering, and statistical testing. While their integration is useful, I am not fully convinced that the advantages of knowledge-boundary fingerprinting over existing approaches are sufficiently established. I would encourage the authors to further clarify the unique contributions and necessity of their method.

  3. Scientific insight and limitations. Although the empirical results are promising, I would appreciate a deeper understanding of why knowledge-boundary probes are particularly effective and under what conditions their effectiveness may be limited. Further discussion of factors such as model updates, correlated responses, and adaptive adversaries could help strengthen the analysis and clarify the practical applicability of the approach.

Questions for authors
  1. Can the authors provide stronger evidence that knowledge-boundary probes offer an advantage over generic reference-stable, contrastively selected factual questions? In particular, how much of the observed discrimination is attributable to the knowledge-boundary hypothesis itself?

  2. How sensitive are the reported false-positive rates and statistical significance levels to correlations among probes? Would a batch-level permutation or bootstrap procedure produce similar conclusions for the primary auditing experiments?

  3. How would KBF distinguish an actual model substitution from a legitimate model version update that changes factual recall behavior? How frequently would reference fingerprints need to be regenerated in such scenarios?

  4. How robust would the method remain against an adaptive relay operator that attempts to identify auditing requests or selectively forwards knowledge-intensive queries to the genuine reference model?

2 comments
AuthorsOpenReview ↗

Response on Practicality, Novelty, False-Positive Control, Reference Drift, and Adaptive Attacks

We sincerely thank the reviewer for the constructive feedback. To address these concerns, we performed further empirical analysis and clarified these situations. Please find our detailed responses below.

1. Practicality and necessity of KBF when users can feel or evaluate model quality downgrading.

KBF provides an evidence-based auditing method that ordinary users can run at very low token costs. The auditing result is highly reliable and can detect mixed partial routing even when users do not observe quality degradation from casual inspection of model responses. Moreover, it provides the capability of long-term and real-time monitoring. Based on real-time monitoring, we found that for some ChatGPT accounts, requests explicitly selecting a premium GPT model such as Astra were instead routed to Luna 5.6. This observation highlights the need for reliable, long-term monitoring.

2. Novelty of KBF method and the scientific insight, compared to generic factual question probes with proper screening

The general idea of using factual probes to detect model substitution is valid, and KBF generally follows this framework. The core scientific contribution of KBF is that the knowledge boundary insight provides a method to “efficiently find an abundant number of probes with high discriminative power.” We have provided empirical evidence of KBF’s efficiency compared to generic factual question generation in Appendix H. KBF generation obtained 1.71–3.45 times as many probes as common-fact generation (Table 1), supporting the efficiency gain. By reviewing of probes from four models and archived audit responses, we found that probes at the knowledge boundary accounted for 66.3%–93.3% of cross-model mismatches in the illustrated pairs (Table 2). This supports a key contribution of KBF, which uses stable responses at the boundaries as model identity signals. We will provide more detailed experiments demonstrating this efficiency gain in the revised paper.

Table 1. Retained probes under the same 60-round generation setting. Both approaches use stability and contrastive screening.

Reference modelKBF specialist generationCommon-fact generation
Sonnet 4.6351137
GPT-5.4338135
DeepSeek V3.230789
GPT-4.1-mini191112

Table 2. Contribution to cross-model mismatches. Each cell shows category mismatches / total mismatches for that model pair (percentage). Stable incorrect includes factual errors and/or equivocal answers.

Reference → TargetStable correctStable incorrect
Sonnet 4.6 → GPT-5.48/44 (18.2%)36/44 (81.8%)
GPT-5.4 → Sonnet 4.634/101 (33.7%)67/101 (66.3%)
Sonnet 4.6 → DeepSeek V3.210/46 (21.7%)36/46 (78.3%)
GPT-4.1-mini → DeepSeek V3.23/45 (6.7%)42/45 (93.3%)

3. Can within-batch probe correlations affect false-positive rates?

This is indeed a valid concern. Although our main approach does not explicitly model probe correlation, it provides a tolerance margin through CP99. We reject only when the one-sided binomial p-value is below 0.05.

Our new FP control experiments yielded 0/60 FPs for each of six models (Table 3), with probe correlations retained. Assuming independent, stable audit rounds, this gives a per-model one-sided 95% FPR upper bound of 4.87%, supporting FP control under the tested conditions. Including the main and September controls, we observed 0/404 FPs in standard audits, plus 0/30 FPs under different wrappers. We will add these analyses to the revision.

Batch bootstrap. We also evaluated a batch-bootstrap alternative, which reduced TP from 155/155 to 154/155 while FP remained 0/16. This suggests that batch resampling may lose detection power by underusing the distinguishing information contributed by individual probes.

Table 3. FP control with fixed probes and calibration. Each model underwent 60 full same-model audits in two collections of 30 rounds.

ModelFP / audits
Claude Sonnet 4.60/60
GPT-5.40/60
GPT-4.1-mini0/60
DeepSeek-V4-Pro-08130/60
Gemini 3 Flash0/60
Qwen3.5-397B-A17B0/60
AuthorsOpenReview ↗

Response on Practicality, Novelty, Probe Correlation, Reference Drift, and Adaptive Attacks (Continued)

4. What if the official APIs get updated and the self-error rates change?

We agree that reference self-error drift is a real concern. KBF assumes a trusted reference backend, but commercial services might change the model behind an endpoint [1][2]. Indeed, the best way to address this issue is to build the reference fingerprint as recently as possible and bind the test result to the fingerprint generation date. That is, we should state the conclusion as “This endpoint is consistent/inconsistent with the fingerprints built on X API at time Y.”

Furthermore, we have empirically shown that most endpoints remained stable (even under silent updates), with calibration allowing for variation. As reported in Appendix D (R5), all 16 models remained consistent through 7 weeks, 14/16 remained consistent at week 9, and 10/11 sampled endpoints remained consistent after approximately 26 weeks. We will clarify that reference endpoints should be retested periodically and fingerprints refreshed when their behavior changes.

5. What if the relay adaptively detects and routes KBF queries to the claimed model while routing other traffic to a cheaper model?

This is a valid concern. Our empirical evaluations provide evidence that KBF can remain effective against basic adaptive routing strategies.

**a) KBF can work without a fixed format, reducing recognizable patterns. Across entity cloze, numerical QA, natural batch QA, and single-question QA, each format achieved TP 10/10 and FP 0/5 across five models (Section 5.4 and Appendix B).

b) Users can mix KBF probes with ordinary queries. We present two approaches to mixing probes and report preliminary experiments on DeepSeek V4 Pro and V4.1 Flash. The main takeaways are: (i) Mixing probes can evade basic attacker detection. When adding a two-turn ordinary conversation as context for each KBF probe batch, all probes passed detectors trained to classify knowledge probing against normal queries. (ii) Mixing probes forces the adaptive router to route user queries to the claimed model, turning adaptive forwarding behavior to the user’s advantage.

Experimental details. (i) We added a two-turn ordinary conversation as context for each single-question QA probe. All 228/228 Pro and 186/186 Flash English probes passed both fixed full-query TF-IDF/logistic-regression detectors trained to distinguish probe-containing queries from normal queries using WildChat [3] and LMSYS-Chat-1M [4]. (ii) We embedded a batch of three probes in each user query and collected their answers from the model response. With each mixing approach, KBF yielded 0/2 FP and 2/2 TP, maintaining detection accuracy. Overall, our preliminary evaluation suggests that KBF offers a more comprehensive and effective defense against adaptive attacks than the baselines discussed in the paper. We will include this analysis in the revision.

[1] DeepSeek. DeepSeek-V3-0324 Release. March 25, 2025.

[2] Google. Continuing to bring you our latest models, with an improved Gemini 2.5 Flash and Flash-Lite release. September 25, 2025.

[3] Wenting Zhao et al. WildChat: 1M ChatGPT Interaction Logs in the Wild. ICLR, 2024.

[4] Lianmin Zheng et al. LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset. ICLR, 2024.

Reviewer_gznPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces KBF, a black-box auditing method that detects whether a API reseller secretly serves a cheaper model.

Critical strengths
  • The problem is real and important, and the cost is low and affordable.
Critical weaknesses
  • It requires special probing questions, and the reseller may detect these patterns. The partial routing experiment uses random routing, but a real reseller may route requests by content, user, or whether it detects the probing.
  • In Figure 7, Sonnet and Opus are close: the mismatch is only 11%. If this is combined with partial routing, it would be hard to distinguish them.
  • As the authors mention themselves, some models' self-error rates can change, so the method needs to be recalibrated.
Questions for authors

When a model's self-error rate changes, does that mean the official provider also change its model? If the method can flag the official endpoint itself as inconsistent, how can it be trusted to detect whether the reseller secretly replaces the model?

3 comments
AuthorsOpenReview ↗

Response on Adaptive Routing, Partial Substitution, and Reference Drift

We sincerely thank the reviewer for the constructive feedback. To address these concerns, we performed further empirical analysis and clarified these situations. Please find our detailed responses below.

1. What if the relay adaptively detects and routes KBF queries to the claimed model while routing other traffic to a cheaper model?

This is a valid concern. Our empirical evaluations provide evidence that KBF can remain effective against adaptive routing strategies under our evaluation setting.

a) KBF can work without a fixed format, reducing recognizable patterns. Across entity cloze, numerical QA, natural batch QA, and single-question QA, each format achieved TP 10/10 and FP 0/5 across five models (Section 5.4 and Appendix B).

b) Users can mix KBF probes with ordinary queries. We present two approaches to mixing probes and report preliminary experiments on DeepSeek V4 Pro and V4.1 Flash. The main takeaways are: (i) Mixing probes can evade basic attacker detection. When adding a two-turn ordinary conversation as context for each KBF probe batch, all probes passed detectors trained to classify knowledge probing against normal queries. (ii) Mixing probes forces the adaptive router to route user queries to the claimed model, turning adaptive forwarding behavior to the user’s advantage.

Experimental details. (i) We added a two-turn ordinary conversation as context for each single-question QA probe. All 228/228 Pro and 186/186 Flash English probes passed both fixed full-query TF-IDF/logistic-regression detectors trained to distinguish probe-containing queries from normal queries using WildChat [1] and LMSYS-Chat-1M [2]. (ii) We embedded a batch of three probes in each user query and collected their answers from the model response. With each mixing approach, KBF yielded 0/2 FP and 2/2 TP, maintaining detection accuracy. Overall, our preliminary evaluation suggests that KBF offers a more comprehensive and effective defense against adaptive attacks than the baselines discussed in the paper. We will include this analysis in the revision.

2. Sonnet and Opus are close. Is the mismatch only 11%?

This appears to be a misunderstanding. The 11% error rate comes from using Opus 4.6 to answer Sonnet 4.6 KBF probes. On the other hand, when using Sonnet 4.6 to answer Opus 4.6 probes (the economically incentivized scenario in practice), the mismatch rate is 22.03%. In fact, our method (Section 5.5) yields a 96.37% detection rate even when the partial routing ratio to Sonnet 4.6 is as low as 26%.

3. What if the official APIs get updated and the self-error rates change?

We agree that reference self-error drift is a real concern. KBF assumes a trusted reference backend, but commercial services might change the model behind an endpoint [3][4]. Indeed, the best way to address this issue is to build the reference fingerprint as recently as possible and bind the test result to the fingerprint generation date. That is, we should state the conclusion as “This endpoint is consistent/inconsistent with the fingerprints built on X API at time Y.”

Furthermore, we have empirically shown that most endpoints remained stable (even under silent updates), with calibration allowing for variation. As reported in Appendix D (R5), all 16 models remained consistent through 7 weeks, 14/16 remained consistent at week 9, and 10/11 sampled endpoints remained consistent after approximately 26 weeks. We will clarify that reference endpoints should be retested periodically and fingerprints refreshed when their behavior changes.

[1] Wenting Zhao et al. WildChat: 1M ChatGPT Interaction Logs in the Wild. ICLR, 2024.

[2] Lianmin Zheng et al. LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset. ICLR, 2024.

[3] DeepSeek. DeepSeek-V3-0324 Release. March 25, 2025.

[4] Google. Continuing to bring you our latest models, with an improved Gemini 2.5 Flash and Flash-Lite release. September 25, 2025.

Thanks for the response. I've raised my score.

Thank you very much for raising your score and for your valuable feedback. We will include these results in the next version.

Reviewer_r336OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes KBF, a black-box auditing method designed to address potential model substitution issues in third-party large language model relay APIs. The core idea is to construct behavioral fingerprints based on the stable responses of a reference model, regardless of whether those responses are correct or incorrect. After identifying stable and discriminative questions, the method determines whether the endpoint under audit deviates significantly from the reference behavior by analyzing fluctuations in the reference model’s own responses. The paper conducted controlled experiments covering eight model families and 16 model endpoints. In the experiments, KBF detected all 155 economically significant substitution directions. The authors further evaluated the method’s robustness and the roles of its individual components through deployment configuration changes, question format variations, and ablation experiments, and identified services on real relay platforms that exhibited statistical inconsistencies with the reference endpoints.

Critical strengths

The research question presented in this paper has clear practical significance. It is difficult to verify, through black-box access, what models are provided by third-party relay APIs. Model substitution could directly impact costs and service quality for users. Addressing this practical need, this paper designs an auditing method that does not require internal model information or cooperation from the provider. It has a clear application objective.

The method described in this paper is designed to account for variations in real-world deployments. Instead of directly treating differences in answers as model substitution, KBF incorporates the inherent fluctuations of the reference model into its assessment through stability screening and reference self-calibration. The ability to regenerate probes and the variety of question formats further enhance the method’s flexibility in continuous auditing. The method also demonstrates good economic feasibility. The paper reports the costs of reference fingerprint construction and online auditing separately, showing that the overhead of repeated audits is low after fingerprint reuse, giving the method potential for practical deployment and periodic detection.

The experiments in this paper are comprehensive, and the controlled results are persuasive. The paper not only evaluates substitution detection across models and within the same model family but also examines deployment configurations, probe formats, mixed routing, and real-world relay platforms, while analyzing the roles of each component through ablation experiments. Together, these experiments support the validity and practical potential of the proposed signal.

In my view, the research on third-party API platforms conducted by the authors is very meaningful, and I consider this to be a topic worthy of further study.

Critical weaknesses

The handling of invalid responses and the audit coverage rate in this paper remain unclear. This paper excludes missing, unparsable, or out-of-scope responses from its statistics, but does not explicitly state the minimum number of valid responses or the coverage rate requirement. According to Algorithm 2, failure to reject the null hypothesis results in a “SAME” return, which may conflate between “the endpoint is consistent with the reference” and “insufficient valid evidence.” We recommend that the authors clarify whether corresponding safeguards are in place in the implementation, report the coverage of valid responses for each experiment, and describe how decisions are made when coverage is insufficient.

There is a certain discrepancy between the assumptions underlying the statistical tests in this paper and the actual request structure. The binomial test used in this paper assumes that probe results are independent of one another, but actual requests may contain up to ten questions from the same domain, and the sharing of context and service state may introduce correlation. The authors have acknowledged this assumption, but its actual impact on significance testing remains unclear. This issue does not invalidate the detection results reported in the paper, but it does limit the interpretation of statistical confidence and false positive control. If the authors could supplement their analysis with a request-batch-based approach and verify whether the main conclusions remain consistent, this would further enhance the credibility of the results.

The ablation studies support the benefits of specialist candidate generation, stability filtering, and contrastive screening. However, the specific contribution of stable incorrect answers remains less directly established. In particular, the paper does not clearly quantify their prevalence in the retained probe sets or their incremental discriminative value relative to stable correct answers. Such an analysis would strengthen the explanation of the proposed fingerprinting signal, although it is not necessary to establish the empirical usefulness of the overall method.

Questions for authors
  1. What should be done if there are not enough valid responses? Is there a minimum number of valid responses or a coverage rate set in the implementation?

  2. Could the authors please indicate whether they have attempted to reanalyze the results of the main experiment and the real-platform audit using request batches as the statistical unit? If request batches are used as the statistical unit, or if a resampling method that preserves the batch structure is employed, would the results for “main substitutions” and false positives for the same reference endpoint remain consistent?

  3. To better understand the contribution of stable incorrect answers, could the authors report the proportion of stable incorrect answers in a fact-checked subset of probes and compare the detection performance when using only stable correct answers, only stable incorrect answers, and a combination of both?

4 comments
AuthorsOpenReview ↗

Response on Valid-Response Thresholds, False-Positive Control, and Stable Incorrect Answers

We sincerely thank the reviewer for the constructive feedback. To address these concerns, we performed further empirical analysis and clarified these situations. Please find our detailed responses below.

1. What should be done if there are not enough valid responses?

We currently target 100–400 probes during generation to balance detection power and cost. More probes can be generated by raising the soft upper limit or adding generation rounds. We handle invalid responses in two ways: (i) For request failures, we retry network and rate-limit errors and return undetermined if target coverage falls below 50%. (ii)For unparsable answers, we will hold the calibrated tolerance fixed and compare decisions when all invalid target responses are counted as matches versus mismatches. If the decisions differ under these extreme assumptions, we recommend repeating the audit, though this does not imply the original result is invalid. In our existing data, approximately 99% of audits did not trigger a recommendation to repeat the audit. The only two flagged cases passed the reliability check in repeat audits.

2. Can within-batch probe correlations affect false-positive rates?

This is indeed a valid concern. Although our main approach does not explicitly model probe correlation, it provides a tolerance margin through CP99. We reject only when the one-sided binomial p-value is below 0.05.

Our new FP control experiments yielded 0/60 FPs for each of six models (Table 1), with probe correlations retained. This gives a per-model one-sided 95% FPR upper bound of 4.87%, supporting FP control under the tested conditions. Including the main and September controls, we observed 0/404 FPs in standard audits, plus 0/30 FPs under different wrappers. We will add this analysis to the revision.

Batch bootstrap. We also evaluated a batch-bootstrap alternative, which reduced TP from 155/155 to 154/155 while FP remained 0/16. This suggests that batch resampling may lose detection power by underusing the distinguishing information contributed by individual probes.

Table 1. FP control with fixed probes and calibration. Each model underwent 60 full same-model audits in two collections of 30 rounds.

ModelFP / audits
Claude Sonnet 4.60/60
GPT-5.40/60
GPT-4.1-mini0/60
DeepSeek-V4-Pro-08130/60
Gemini 3 Flash0/60
Qwen3.5-397B-A17B0/60

3. Effectiveness of Stable Incorrect Answers

We reviewed probes from four models and classified them as stable correct or stable incorrect (Table 2). The latter includes factual errors and equivocal answers arising from uncertainty about the facts or ambiguity in question interpretation, which remain stable for a given model. Stable incorrect probes made up 49.79% of the reviewed set but contributed 66.3%–93.3% of all mismatches in the four illustrated model pairs (Table 3), providing most of the observed model discrimination. We will include these results in the revision to further clarify the contribution of the knowledge boundary.

Table 2. Review of 1,187 specialist probes from four models.

ModelTotalStable correctStable incorrect
Sonnet 4.6351198153 (43.59%)
GPT-5.4338175163 (48.22%)
DeepSeek V3.2307149158 (51.47%)
GPT-4.1-mini19174117 (61.26%)
Total1,187596591 (49.79%)

Table 3. Contribution to cross-model mismatches. Each cell shows category mismatches / total mismatches for that model pair (percentage).

Reference → TargetStable correctStable incorrect
Sonnet 4.6 → GPT-5.48/44 (18.2%)36/44 (81.8%)
GPT-5.4 → Sonnet 4.634/101 (33.7%)67/101 (66.3%)
Sonnet 4.6 → DeepSeek V3.210/46 (21.7%)36/46 (78.3%)
GPT-4.1-mini → DeepSeek V3.23/45 (6.7%)42/45 (93.3%)

Thank you for the additional experiments and clarifications. The proposed sensitivity check addresses my concern about unparsable responses, and the probe level analysis helps clarify how stable incorrect answers contribute to model discrimination.

My main remaining concern is false-positive control under correlated probes, though the additional same model audits are reassuring for the settings tested.

Overall, the rebuttal strengthens my confidence in the method, and I maintain my rating of 3.

Thank you very much for your valuable feedback. We will include these results in the next version.

Thank you for your continued engagement and increased confidence in our work. Having reflected further on your remaining concern, we would like to offer a brief clarification.

In the paper, we use a binomial approximation under an independence assumption with a nominal significance level of 5%. Our additional experiments directly test whether the full procedure preserves this intended false-positive control when real-world probe correlations are retained. Because each complete audit is treated as a single observation, this empirical validation requires independent, stable audit rounds rather than independence across probes within an audit. The one-sided 95% upper confidence bound on the false-positive rate is below 5% for each tested model. These empirical results complement our assumption-based analysis by showing that the procedure maintains low false-positive rates with the probe correlations present in the tested settings.

Reviewer_2NMSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces KBF, a method for checking whether an LLM API is serving the model it claims to serve. The method uses factual questions that a particular model answers consistently (including consistently wrong answers), to distinguish it from other models. Experiments across 16 models show strong detection performance, and the authors also test the method on real API providers.

Critical strengths

(1) First, the idea of using stable factual recall, including consistently wrong answers, as a model fingerprint, is an interesting idea. It seems to offer a practical way to audit models without needing access to their weights, logits, etc.

(2) The experiments are fairly thorough, covering multiple models, several baseline methods, different deployment settings, and APIs. The authors also provide useful ablations and robustness analyses.

Critical weaknesses

(1) In my opinion, the statistical evaluation could be more rigorous. In particular, the paper should account for dependencies between probes and provide stronger evidence that false-positive rates are well controlled. E.g. questions are grouped into batches and come from related domains, so their responses could correlated, leading to issues with reported p values.

(2) The method is not thoroughly tested against an adaptive adversary method. For example, a dishonest API provider might recognize the audit questions and selectively route them to the correct model. The paper discusses this possibility but does not directly evaluate such an attack.

Questions for authors

Would it be possible for the authors to evaluate the main detection method using a statistical test that accounts for correlations between probes within the same batch?

1 comment
AuthorsOpenReview ↗

Response on False-Positive Control and Adaptive Routing

We sincerely thank the reviewer for the constructive feedback. To address these concerns, we performed further empirical analysis and clarified these situations. Please find our detailed responses below.

1. Can within-batch probe correlations affect false-positive rates?

This is indeed a valid concern. Although our main approach does not explicitly model probe correlation, it provides a tolerance margin through CP99. We reject only when the one-sided binomial p-value is below 0.05. Our new FP control experiments yielded 0/60 FPs for each of six models (Table 1), with probe correlations retained. This gives a per-model one-sided 95% FPR upper bound of 4.87%, supporting FP control under the tested conditions. Including the main and September controls, we observed 0/404 FPs in standard audits, plus 0/30 FPs under different wrappers. We will add these analyses to the revised manuscript.

Batch bootstrap. We also evaluated a batch-bootstrap alternative, which reduced TP from 155/155 to 154/155 while FP remained 0/16. This suggests that batch resampling may lose detection power by underusing the distinguishing information contributed by individual probes.

Table 1. FP control with fixed probes and calibration. Each model underwent 60 full same-model audits in two collections of 30 rounds.

ModelFP / audits
Claude Sonnet 4.60/60
GPT-5.40/60
GPT-4.1-mini0/60
DeepSeek-V4-Pro-08130/60
Gemini 3 Flash0/60
Qwen3.5-397B-A17B0/60

2. What if the relay adaptively detects and routes KBF queries to the claimed model while routing other traffic to a cheaper model?

This is a valid concern. Our empirical evaluations provide evidence that KBF can remain effective against basic adaptive routing strategies.

a) KBF can work without a fixed format, reducing recognizable patterns. Across entity cloze, numerical QA, natural batch QA, and single-question QA, each format achieved TP 10/10 and FP 0/5 across five models (Section 5.4 and Appendix B).

b) Users can mix KBF probes with ordinary queries. We present two approaches to mixing probes and report preliminary experiments on DeepSeek V4 Pro and V4.1 Flash. The main takeaways are: (i) Mixing probes can evade basic attacker detection. When adding a two-turn ordinary conversation as context for each KBF probe batch, all probes passed detectors trained to classify knowledge probing against normal queries. (ii) Mixing probes forces the adaptive router to route user queries to the claimed model, turning adaptive forwarding behavior to the user’s advantage.

Experimental details. (i) We added a two-turn ordinary conversation as context for each single-question QA probe. All 228/228 Pro and 186/186 Flash English probes passed both fixed full-query TF-IDF/logistic-regression detectors trained to distinguish probe-containing queries from normal queries using WildChat [1] and LMSYS-Chat-1M [2]. (ii) We embedded a batch of three probes in each user query and collected their answers from the model response. With each mixing approach, KBF yielded 0/2 FP and 2/2 TP, maintaining detection accuracy. Overall, our preliminary evaluation suggests that KBF offers a more comprehensive and effective defense against adaptive attacks than the baselines discussed in the paper. We will include this analysis in the revision.

[1] Wenting Zhao et al. WildChat: 1M ChatGPT Interaction Logs in the Wild. ICLR, 2024.

[2] Lianmin Zheng et al. LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset. ICLR, 2024.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for the Area Chairs

Dear Area Chair,

Thank you and the reviewers for the helpful feedback on our submission. We have addressed all the points raised with additional experiments and clarifications. Following our rebuttal, both reviewers who participated in the discussion gave positive feedback; Reviewer gznP raised their score, and Reviewer r336 noted increased confidence in our approach.

Our paper presents KBF, a black-box auditing protocol that uses stable responses near the knowledge boundary as model fingerprints. KBF combines accurate, robust detection with benign, renewable probes and low audit costs. In real-world API use, providers control which backend serves each request, while users cannot directly verify the model’s identity, making independent black-box auditing necessary. Our recent observation of unexpected behavior at OpenAI's official endpoints, detailed in our follow-up comment, shows that there is a broad need for tools like KBF across many real-world deployment settings.

Below, we summarize the additional evidence obtained during the rebuttal and how it supports KBF's core claims and effectiveness in three areas:

(1) Empirical false-positive control: Across repeated full audits, KBF keeps false-positive rates low even with probe correlations present, achieving per-model upper confidence bounds below 5%. These results confirm that our independence-based test holds up well in practice.

(2) Adaptive routing defenses: We proposed two probe-mixing methods that mitigate basic adaptive routing attacks while maintaining detection accuracy.

(3) Knowledge-boundary contribution: In the illustrated model pairs, stable incorrect answers account for 66.3%–93.3% of all cross-model mismatches. These probes provide most of the observed model discrimination, showing that probing the knowledge boundary is key to effective behavioral fingerprinting.

We also clarified how to handle reference drift caused by routine model updates and added an invalid-response sensitivity check that Reviewer r336 acknowledged as addressing their concern about unparsable responses.

We will incorporate all new experimental results, clarified statistical assumptions, and discussions of practical limitations into the next version.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.