acceptodds
Under review as a conference paper at ICLR 2027

Mole: Detecting Insider Threats in AI Agents

Abstract

Model misalignment, prompt injection, or operator misuse could lead AI agents operating frontier-lab accounts to exfiltrate model weights, poison training data, or weaken release gates. Existing benchmarks do not test whether defenders can detect this activity among routine work under a limited review budget. We introduce Mole, an open benchmark of 150 AI-operated accounts sharing 9 stateful services over 30 workdays, with 12 threats and 8 corpora from four models totaling roughly 20 billion tokens. Of 39 agent models, 72% complete most assigned harmful objectives. Mole enables comparison of 40 monitors across corpus generators, observability levels, and threats; even the best monitor in our single-day audit-event comparison misses nearly half of completed harm. Mole also enables monitor development: benchmark-guided search improves a mid-tier monitor by 49-64%, while selective use of a stronger monitor improves budget-AUC by 10% over applying it to every account-day at comparable modeled cost.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
insider threat detection, AI agents, agent monitoring, AI safety, AI control, benchmark, LLM agents, frontier AI labs, semantic monitoring

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_3UJ5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces MOLE as both a benchmark for insider-threat detection and an environment for developing better monitors. The authors use an automated research agent to search for monitoring strategies. These strategies combine model judgments with statistical features, including unusual actions and action sequences. The results show improvements over a uniform GPT-4.1 monitor across multiple corpora. Also, using a stronger model also improves detection relative to applying that model uniformly at a similar modeled cost.

Critical strengths

The experimentation is fairly broad. I liked that normal and harmful activity in each corpus come from the same generator. the time-based splits and the tests on unseen accounts also make sense to me. The authors also check whether harmful objectives were actually completed rather than treating every assigned attack as successful.

Critical weaknesses

a. The automated search only uses a small validation set. the research agent tries 185 ideas, but it checks them on just three validation days and it can see the overall validation score again and again. this makes it easy to end up with strategies that only work well on those few incidents. b. The human check also covers only a few different examples. there are 540 outcome ratings, but they come from just 32 unique account-days. the raters agree a lot, which is good, but the sample has no refusal cases. so this mainly shows the labels are right for these samples. it doesn’t show they are reliable for every generator, threat and outcome type. c. it’s also not shown that this holds up against attackers who adapt. the stealthy prompts tell agents to blend into normal work, but the attackers never get feedback from a specific monitor. so I still don’t know how well detection works when attackers actively learn to get past the defender.

Questions for authors
  1. How well do the monitors transfer to a different organization configuration or a held-out threat family?
  2. Do the main rankings and threat-specific findings remain stable across independent simulation runs?
  3. Could you report detection before harm including time to the first alert and campaign-level recall?
  4. How does monitor performance change when attackers can adapt using feedback from the defender?
Reviewer_UADfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces MOLE, a benchmark for detecting insider threats among AI-operated accounts in a simulated frontier AI lab. The benchmark consists of 150 persistent accounts interacting with nine shared, stateful services over 30 workdays, with twelve threat scenarios spanning conventional insider threats and ML-specific attacks. MOLE studies whether a monitor can identify accounts that have completed harmful objectives among a larger population performing routine work, under a limited daily investigation budget. The authors generate eight corpora using four different agent models and evaluate 39 models on harmful-objective execution and 40 monitors in their primary audit-event comparison. They investigate how detection varies with monitor capability, observability, corpus generator, and threat type. The paper also demonstrates how benchmark feedback can guide the development of monitoring strategies, including statistical feature augmentation and selective use of stronger LLM monitors.

Critical strengths
  1. Potentially useful benchmark: After reviewing the literature, I believe MOLE brings together several aspects of insider-threat detection and agent evaluation that have previously been studied separately. The combination of persistent AI-operated accounts, shared stateful services, routine organizational activity, and daily budget-constrained monitoring seems like a meaningful contribution.

  2. Extensive evaluation: Evaluating multiple corpus generators and monitor models allows the authors to investigate not just aggregate performance, but also differences across observability levels and threat types. I found the observation that additional context can sometimes outweigh monitor capability particularly interesting. I also appreciate RQ5, where the authors demonstrate that the benchmark can guide monitor development rather than simply compare fixed approaches.

Critical weaknesses
  1. The comparison between semantic and classical monitors may not fully isolate the claimed advantage of semantic monitoring. The semantic monitors receive relatively rich, ordered service-call traces, including resource information and call arguments, while the classical baselines generally operate on aggregate statistical features or restricted resource-access representations. This makes me wonder how much of the observed performance gap comes from semantic reasoning itself, rather than differences in the information and representations available to the detectors.

The supervised baseline with additional statistical features in Appendix K.2 is a useful control, and I appreciate that the authors also tune their FACADE[4] implementation. However, prior insider-threat detection work includes more expressive sequence-aware and graph-based approaches, such as ADSAGE[2] and LAN[3], which seem relevant here.

I believe the results convincingly establish an advantage over the classical implementations evaluated here, but I am less certain whether this gap would persist against more expressive sequence-aware or graph-based approaches using comparable information. I would be interested in hearing the authors' thoughts on this.

  1. I also have a couple of smaller concerns that I would appreciate clarification on. a. First, I may be misunderstanding the labeling protocol, but Appendix B.2 appears to treat a collusive participant's completed contribution as a positive outcome, even if the overall harmful campaign has not necessarily succeeded. This seems somewhat different from the completed-harm criterion emphasized elsewhere in the paper, and I wonder how this distinction affects the interpretation of the monitoring results.

b. Second, while the evaluation covers multiple agent generators and includes useful robustness analyses, it is still conducted within a single synthetic organization with fixed roles, workflows, and threat scenarios. Some threat-level results also rely on relatively few positive examples. I believe these limitations make it difficult to assess how consistently the findings would generalize to different organizational settings or unseen threats. I would be interested in the authors' perspective on these limitations and, where feasible, additional evidence supporting the robustness of the main conclusions.

Questions for authors
  1. Have the authors examined whether benign and harmful account-days differ systematically in properties such as trace length, service usage, or resource types? I would be interested in whether performance remains consistent when controlling for these properties, or when evaluating benign activities involving similarly sensitive resources and operations. This could help clarify whether monitors are exploiting dataset-specific artifacts.

  2. Could the authors elaborate on the baseline selection, particularly the absence of richer sequence-aware or graph-based approaches such as ADSAGE and LAN? I appreciate the supervised baseline in Appendix K.2, but wonder whether stronger temporal representations or longer training periods would narrow the observed gap.

  3. Appendix B.2 seems to label each colluder's completed contribution as positive, even when the overall harmful campaign may not have succeeded. Could the authors clarify how this aligns with the completed-harm criterion in the main text and whether it affects the evaluation?

  4. Does the reciprocal-rank fusion in Appendix F.6 rank account-days within each day or across the entire evaluation period? The latter could introduce future information. Also, Table 9 disables reasoning for some generators, yet Table 14 reports reasoning content for them. Could the authors clarify how reasoning is defined and whether it is comparable across generators?

[1] Yu, J., Xie, X., Hu, Q., Ma, Y., & Zhao, Z. (2026). Chimera: Harnessing Multi-Agent LLMs for Automatic Insider Threat Simulation. [2] Garchery, M., & Granitzer, M. (2020). ADSAGE: Anomaly Detection in Sequences of Attributed Graph Edges applied to insider threat detection at fine-grained level. [3] Cai, X., Wang, Y., Xu, S., Li, H., Zhang, Y., Liu, Z., & Yuan, X. (2024). LAN: Learning Adaptive Neighbors for Real-Time Insider Threat Detection. [4] Kantchelian, A., Neo, C., Stevens, R., et al. (2026). Facade: High-Precision Insider Threat Detection Using Deep Contextual Anomaly Detection.

Reviewer_AgSNOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary
  1. This paper presents Mole, a benchmark of AI agents operating accounts in a simulated frontier AI lab. The benchmark includes 12 threats mapped to MITRE ATT&CK and ATLAS, and 8 corpora (4 generator models, each with a single-day and a multiday campaign). Most of the 39 agent models tested complete the majority of their assigned harmful objectives.
  2. Evaluations on Mole show that semantic monitors outperform classical baselines at identifying threats, and that observability can matter more than monitor strength: a weaker monitor given agent reasoning can outperform a stronger monitor restricted to audit events. Even the best monitor misses nearly half of completed harm in the single-day audit-event comparison.
  3. They show that Mole can be used to develop better monitors through an automated research experiment, in which a monitor optimised against Mole is more effective than the baselines at the same or lower modelled cost.
Critical strengths
  1. Mole is a useful contribution to developing monitors that detect threats in a frontier lab setting. The benchmark covers complex, multi-day activity using realistic tools. It will allow groups outside frontier labs to run control and monitoring research at a larger scale and in a setting closer to real deployments, which I would expect to lead to monitors that work better in real-world settings.
  2. The demonstration of how Mole can be applied to develop improved monitors is valuable. It supports the value of Mole as a benchmark and shows a possible route forward for using automated research agents to develop more robust safeguards efficiently.
Critical weaknesses
  1. The novelty of Mole relative to prior work could be clearer. The related work section in the appendix (particularly Table 2) helps address this, but readers would benefit from a concise version in the main body that explains Mole’s novel contributions. I appreciate that this was likely left out of the main body due to the page limit.
  2. Minor weakness: some results are hard to read and verify. In particular, Figure 8(b) has very small text labels that overlap with each other and with the frontier line, making it hard to interpret. The appendix also contains a large amount of detail, and results that affect the main claims are hard to find. The appendix may benefit from being reviewed, with unnecessary detail cut or moved to supplementary materials with the Mole release, so that readers can easily find the experimental details and ablations that influence the main results.
Questions for authors
  1. Can you more clearly frame, in the main body, how Mole differs from similar benchmarks (as shown in Table 2)? In particular, what does Mole let researchers measure that these benchmarks do not, and how does that make it more or differently useful for developing better monitors?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and reviewer discussion

We thank the reviewers and Area Chair for their time and consideration. As the discussion period concludes, we wanted to briefly summarize the revisions and subsequent reviewer feedback.

In response to the reviews, we substantially expanded the evaluation with stronger classical baselines (ADSAGE, LAN, and supervised TF-IDF), activity-matched comparisons, held-out-threat evaluation, independent-run stability analysis, additional human validation, campaign-level early detection, and a live adaptive attacker–defender pilot. We also strengthened the paper's positioning relative to prior benchmarks and clarified several methodological details.

We appreciate the positive feedback following these additions. Reviewer UADf increased their rating from 2 to 3, noting that the remaining concerns do not substantially detract from the contribution. Reviewer AgSN confirmed satisfaction with the rebuttal while maintaining their rating of 3. Reviewer 3UJ5 also rated the paper 3.

We hope these revisions and the discussion help inform the Area Chair's final assessment, and we thank everyone again for their constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.