Pith.

Artificial Intelligence · arXiv:2610.27115 v1 · submitted 19 Oct 2026 · 78 references

PROVISO: Provenance-Informed Action Selection for Secure LLM Agents

Bo Yan, Qi Zhang, Ziwen Kan, Fnu Suya, Song Wang

The paper says

“LLM agents act on the external world through tools and third-party skills, so a single executed action can cause irreversible harm.”

more of the abstract

Existing defenses judge a proposed action without tracking whether its content comes from the user or from external sources such as tool outputs, and they judge it in isolation rather than against an available alternative, which forces a trade-off between security and utility. We present PROVISO, a controller that decides which action to commit after the agent proposes an action and before it is executed. PROVISO encodes trusted observations, external observations, the proposed action, and an alternative into provenance-typed memory, relates them through role-specific queries, and commits the action with the highest calibrated value learned from outcomes. On four agent security tasks spanning the agent lifecycle, no baseline exceeds PROVISO in both security and utility. PROVISO improves F1 over the strongest baseline by 13.84 points on SkillsBench and by 7.07 points on ToolSafe, and it reaches 88-99 F1 on three additional skill detection benchmarks.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 21 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up. Solid engineering contribution with thorough evaluation; the central claim holds on tested benchmarks but assumes a complete trusted contract K, which is a known limitation the authors acknowledge. This line is wrong

2 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 78 references say.

The paper says, at Section 3.1, Limitations (third paragraph of Limitations section)

The review did not quote this passage.

Pith says: Objection 1 of 2 · major

The entire admissibility rule and the controller's learning target depend on the trusted contract K being complete and correctly specified (Section 3.1, 'We say an operation or target is licensed by K when the task contract permits it'). The paper acknowledges that PROVISO is fitted per benchmark and that attacks adapted to it are not evaluated, but it does not discuss the sensitivity to incomplete or ambiguous K.

more of objection 1

In real deployments, user requests may omit necessary constraints or the deployment policy may be underspecified. Under such gaps, PROVISO's learned mapping may fail, while static rules or LLM reasoning might be more robust. The paper should add a discussion (or experiment) on how the decision degrades when K is missing constraints or contains ambiguities, and explicitly state the assumption of K completeness as a limitation.

The paper says, at Section 5.1, Table 2 (ANCHOR agent misuse)

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

The claim that PROVISO 'reduces H&R to zero while completing all benign tasks' on ANCHOR is presented as a headline result, but the benign completion is measured on a fixed panel of 80 tasks. The paper states that AGrail completes 35/80 (43.75%) on the same panel, but it does not report whether the benign tasks are representative or if the 80 tasks are a complete set from the benchmark.

more of objection 2

The authors should clarify whether the benign completion metric is based on the full ANCHOR benign test set or a subset, and whether the comparison with AGrail is statistically significant (e.g., via a paired test).

5 minor objections

The paper says, at Section 3.2, Eq. (3) and (4)

Pith says: minor objection 1 of 5

The notation for role embeddings e_r in Eq. (3) is introduced but the dimension of e_r is not stated. Please specify the dimension (presumably 384, matching the projection width).

The paper says, at Section 4, 'Baselines' paragraph

Pith says: minor objection 2 of 5

The list of baselines for each task is thorough, but it would be helpful to state explicitly which baselines use training data (e.g., fine-tuned models) and which use frozen LLMs or static rules. Table 8 partially addresses this, but the text could cross-reference it more clearly.

show 3 more

The paper says, at Section 5.3, Figure 4 caption

Pith says: minor objection 3 of 5

The caption says 'Paired 95% confidence intervals are in Table 15', but Table 15 is in Appendix F. Please reference the appendix table explicitly (e.g., 'Appendix F, Table 15').

The paper says, at Appendix I, Case I.1

Pith says: minor objection 4 of 5

The case study would be more informative if it showed the input to PROVISO (K, G, P0, P1) explicitly, not just the outcome. The provided text describes the context but the reader must infer the trust/external split.

The paper says, at Throughout

Pith says: minor objection 5 of 5

The paper uses many abbreviations (H&R, ASR, BCE, etc.) that are defined once and then used extensively. A glossary or an abbreviation table in an appendix would help readers.

What would settle it

For each of the four tasks, construct a test set where the trusted observation K is systematically perturbed: (a) remove one required constraint (e.g., omit the authorized recipient from the user request on ToolSafe), (b) rephrase K with underspecified details (e.g., 'pay the bill' without amount or recipient), and (c) add an ambiguous directive that could be interpreted as authorizing multiple operations. Run the same PROVISO and the top three baselines per task. Measure the relative change in F1/utility/ASR. If PROVISO's advantage over the strongest baseline is no longer statistically significant (paired test, p>0.05) on any of the perturbed sets, then the central claim is conditional on complete K and the risk is real.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 21 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 59 s · 30,657 tokens read, 5,018 written
  3. Editor wrote the verdict above 13 s · 31,353 tokens read, 950 written
  4. Referee minor revision: 2 major objections, 5 minor 19 s · 31,012 tokens read, 1,535 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The paper’s core idea—provenance-informed action selection at the commitment point—is a genuinely new framing for LLM agent security. It moves beyond judging actions in isolation by separating trusted user input from external content and learning to compare a proposed action against an alternative. The architecture (typed memory, role-query deliberation, calibrated commitment) is sensible and the ablation study convincingly shows that the provenance tracking and the retrieval structure both matter.

The evaluation is the paper's strongest asset. Four benchmarks covering the agent lifecycle, three additional skill detection benchmarks, statistical tests, and controlled component analysis. The results are clear: on every task, no existing defense beats PROVISO on both security and utility. The margins on SkillsBench and ToolSafe are substantial.

The stress-test concern about the trusted contract K being complete is real but not a hidden flaw. The paper states in Section 3.1 that K defines the contract and that it is taken as given. The limitations section explicitly says the method requires labeled outcomes from the target setting and does not study transfer or imperfect specifications. So the authors are upfront about the scope. What is missing is a sensitivity analysis—how performance degrades when K is incomplete or ambiguous—but that is a reasonable extension, not a fatal omission. The paper does not claim robustness to imperfect K.

The other soft spots are minor: candidate selection requires restorable environments and the cost of two rollouts; adaptive attacks are not evaluated. Again, these are acknowledged.

This is a well-engineered paper for a practical security problem. The results are reproducible (code released) and the evidence supports the claims within the stated scope. It does not reshape a field, but it provides a solid, new approach that practitioners could adopt.

Recommendation: Send to peer review. A reviewer should push for a discussion of K completeness and perhaps some synthetic evaluation, but the paper as it stands is a legitimate contribution.

The strongest objection

The paper's formulation (Section 3.1) defines admissible actions relative to trusted observations K = (O, S, R, T). The controller learns to value actions from outcomes, but K is taken as given. All four benchmarks provide well-specified K; the paper does not evaluate cases where K is incomplete or ambiguous. In real deployments, user requests may omit necessary constraints (e.g., 'pay the invoice' without specifying which account), or the deployment policy may be underspecified. Under such gaps, PROVISO's learned mapping may fail: it could block benign actions that rely on implicit authorization (high false positive) or allow attacks that exploit missing coverage (high false negative). Baselines that use static rules or LLM reasoning might be more robust to missing specification because they rely on general world knowledge. Therefore, the empirical superiority of PROVISO on these four tasks may not transfer to realistic settings where K is imperfect. This is a load-bearing assumption because the entire admissibility rule (and the controller's learning target) depends on K being correct and complete. The paper does not test this sensitivity, leaving the robustness of the central claim unexamined.

The reader's verdict is ACCEPT with high confidence, and the paper's evidence is solid for the given benchmarks. However, the weakest assumption identified by the reader – completeness of K – is indeed the most load-bearing. The paper does not address it experimentally, relying on benchmark-provided K which is fully specified. In practice, users and deployments may provide incomplete or ambiguous contracts, and the controller's performance under those conditions is crucial for the claimed superiority. This concern does not invalidate the presented results, but it limits their generalizability. It does not change the verdict on the paper as presented, because the paper honestly describes its setup and limitations. Therefore, verdict_should_be = UNCHANGED. I agree with the reader's identification of the weakest assumption. The paper's self-stated limitations acknowledge the benchmark-specific nature and lack of transfer, but do not explicitly test K incompleteness. This gap is significant but does not undermine the core contribution.

The referee's report

If the empirical results hold as reported, PROVISO offers a principled solution to a key challenge in LLM agent safety: making action decisions that depend on the provenance of content rather than judging actions in isolation. The unified formulation across admission and candidate selection, together with the learnable architecture, could provide a foundation for practical guardrails. The thorough empirical evaluation across four diverse benchmarks with multiple baselines and component analyses lends credibility to the claims. The paper also releases code, supporting reproducibility.

Pith's reading of the paper (a machine's summary, not the authors')

PROVISO is a controller that decides which action an LLM agent should commit, placed between the agent's policy and the executor. It addresses two deficiencies of existing defenses: they ignore the provenance of action content (whether it comes from the trusted user or from external tool outputs) and they judge actions in isolation without comparing them to a feasible alternative. PROVISO encodes observations by provenance type, uses role-specific queries to deliberate over them, and learns the value of each candidate action from outcomes, then calibrates an operating point. Across four agent security tasks—malicious skill detection, agent misuse, unsafe tool invocation, and indirect prompt injection—no published baseline surpasses PROVISO in both security and utility. For example, PROVISO improves F1 by 13.84 points on SkillsBench and by 7.07 points on ToolSafe over the strongest baseline.

What carries the argument. Three components: (i) provenance-typed memory, which projects each observation segment into a learned role embedding (K, G, P0, or P1) so that identical text from different sources produces distinct representations; (ii) role-query deliberation, where five learned query states (four role queries plus a commitment state) pass through L blocks of cross-attention, self-attention, and feed-forward updates to retrieve, compare, and integrate evidence; (iii) calibrated commitment, which estimates action value from the deliberation states and applies per-action offsets selected on development data to choose the best candidate.

Core claim. PROVISO introduces a unified formulation for agent security decisions at the commitment point: separate trusted observations (K, from user and deployment) from external observations (G, from tool outputs and skills), and choose between the proposed action P0 and an alternative P1 by comparing learned calibrated values. Its architecture—provenance-typed memory, role-query deliberation with five queries that retrieve and compare across memory, and calibrated commitment trained on native outcomes—achieves both high security and high utility across benchmarks spanning skill installation to indirect prompt injection. The paper establishes that provenance-agnostic judgment and isolation without alternatives are the root causes of the security-utility trade-off in existing defenses.

The premise it rests on. The trusted contract K (the user request and deployment scope) is assumed to be complete and correctly specified; if K is incomplete or ambiguous, PROVISO may block benign actions that rely on implicit authorization or fail to detect attacks that exploit missing specifications.

If the paper is right (Pith's inferences, not the authors')

PROVISO can be deployed in any LLM agent pipeline that has a commitment point, from skill installation to final action execution, replacing handcrafted rules with learned value comparison.

The unified formulation reduces the need for separate defenses for different attack stages; the same controller architecture handles skill detection, misuse, unsafe calls, and indirect injection with only per-deployment weight fitting.

Because PROVISO compares the proposed action with an alternative (refusal or recovery), it avoids the forced trade-off between blocking attacks and completing benign tasks that plagues threshold-based guards.

On the four agent-security benchmarks, PROVISO sets new state-of-the-art results: zero Harm & Risk score on ANCHOR, 99.12 F1 on ToolSafe, 85.81 F1 on SkillsBench, and attack success rate of only 0.28% on AgentDojo while maintaining high utility.

PROVISO requires labeled outcome data for each deployment and, in candidate-selection mode, sandboxed rollouts of both candidates before commitment. This limits its applicability in environments where rollouts are expensive or unavailable, a limitation the authors acknowledge.

The paper evaluates only attacks present in released benchmarks; adaptive attacks designed specifically to bypass PROVISO's provenance-based value estimation are not studied, leaving open how robust the method would be against adversaries who craft content to manipulate the learned queries.

If PROVISO's provenance-typed memory and calibrated commitment prove effective, the same pattern—separating trusted from external sources, reasoning over alternatives, and tuning an operating point from data—could generalize to other domains where an agent must choose among actions with different authority and risk profiles.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

LLM agents act on the external world through tools and third-party skills, so a single executed action can cause irreversible harm. Existing defenses judge a proposed action without tracking whether its content comes from the user or from external sources such as tool outputs, and they judge it in isolation rather than against an available alternative, which forces a trade-off between security and utility. We present PROVISO, a controller that decides which action to commit after the agent proposes an action and before it is executed. PROVISO encodes trusted observations, external observations, the proposed action, and an alternative into provenance-typed memory, relates them through role-specific queries, and commits the action with the highest calibrated value learned from outcomes. On four agent security tasks spanning the agent lifecycle, no baseline exceeds PROVISO in both security and utility. PROVISO improves F1 over the strongest baseline by 13.84 points on SkillsBench and by 7.07 points on ToolSafe, and it reaches 88-99 F1 on three additional skill detection benchmarks.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-21

The authors' main claim, as the classifier picked it out 0.99

“On four agent security tasks spanning the agent lifecycle, no baseline exceeds PROVISO in both security and utility.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.10 · states limitation 0.07 · human participants 0.06 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
LLM agents act on the external world through tools and third-party skills, so a single executed action can cause irreversible harm. background 1.00 0.05 0.02 asserted 0.99
Existing defenses judge a proposed action without tracking whether its content comes from the user or from external sources such as tool outputs, and they judge it in isolation rather than against an available alternative, which forces a trade-off between security and utility. gap 0.97 0.04 0.05 asserted 1.00
We present PROVISO, a controller that decides which action to commit after the agent proposes an action and before it is executed. contribution 0.96 0.85 0.07 asserted 1.00
PROVISO encodes trusted observations, external observations, the proposed action, and an alternative into provenance-typed memory, relates them through role-specific queries, and commits the action with the highest calibrated value learned from outcomes. method 0.99 0.89 0.09 asserted 1.00
On four agent security tasks spanning the agent lifecycle, no baseline exceeds PROVISO in both security and utility. result 1.00 0.94 0.66 asserted 1.00
PROVISO improves F1 over the strongest baseline by 13.84 points on SkillsBench and by 7.07 points on ToolSafe, and it reaches 88-99 F1 on three additional skill detection benchmarks. result 1.00 0.96 0.21 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-21: 1 request, 5,036 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (6)
Figure 1 of the paper
Figure 1. A payment task with an injected instruction.
Figure 2 of the paper
Figure 2. Overview of PROVISO. Left: at step t, the agent’s observations are split by provenance into trusted (K) and external (G) ones; the LLM policy proposes P0, and an alternative P1 is available. Middle: PROVISO builds role-typed memory from x = (K, G, P0, P1), lets four role queries and a commitment state (C) retrieve, compare, and update over L blocks, then values each candidate and commits a ∗ . Right: the executor runs a ∗ , and the next observation joins G. Dashed arrows show training: native outcomes fit the parameters, and the offsets τ are calibrated on development data. block…
Figure 3 of the paper
Figure 3. Security and utility of PROVISO and baselines on four agent security tasks. The upper right is better.
Figure 4 of the paper
Figure 4. Effect of changing one component of PRO
Figure 5 of the paper
Figure 5. TOOLSAFE results across (a) twenty training runs, (b) proposer backbones, and (c) logit-threshold offsets for the selected model; offset zero is the selected threshold. −5 0 5 0 20 40 60 80 100 Risk-logit threshold offset Percent ToolSafe −5 0 5 0 20 40 60 80 100 Risk-logit threshold offset Percent ANCHOR −5 0 5 0 20 40 60 80 100 Risk-logit threshold offset Percent SkillsBench F1 Unsafe recall Benign retention
Figure 6 of the paper
Figure 6. Effect of shifting the decision threshold on the logit scale for

Figures are the authors', reproduced from the paper.

References (78)
  1. [1] Debenedetti, Edoardo and Zhang, Jie and Balunovic, Mislav and Beurer-Kellner, Luca and Fischer, Marc and Tram. Advances in Neural Information Processing Systems 37 , pages =. 2024 , doi = on Pith
  2. [2] Li, Hao and Liu, Xiaogeng and Chiu, Chun and Li, Dianqi and Zhang, Ning and Xiao, Chaowei , booktitle =. 2025 , doi = on Pith
  3. [3] An, Hengyu and Zhang, Jinghuai and Du, Tianyu and Zhou, Chunyi and Li, Qingming and Lin, Tao and Ji, Shouling , booktitle =. 2025 , doi = on Pith
  4. [4] Li, Hao and Liu, Xiaogeng and Zhang, Ning and Xiao, Chaowei , booktitle =. 2025 , doi = on Pith
  5. [5] Zhu, Kaijie and Yang, Xianjun and Wang, Jindong and Guo, Wenbo and Wang, William Yang , booktitle =. 2025 , publisher = on Pith
  6. [6] Benchmarking and Defending Against Indirect Prompt Injection Attacks on Large Language Models , author =. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year = on Pith
  7. [7] Chen, Sizhe and Piet, Julien and Sitawarin, Chawin and Wagner, David , booktitle =. 2025 , publisher = on Pith
  8. [8] Chen, Sizhe and Zharmagambetov, Arman and Mahloujifar, Saeed and Chaudhuri, Kamalika and Wagner, David and Guo, Chuan , booktitle =. 2025 , doi = on Pith
  9. [9] Zhan, Qiusi and Liang, Zhixiang and Ying, Zifan and Kang, Daniel , booktitle =. 2024 , doi = on Pith
  10. [10] Greshake, Kai and Abdelnabi, Sahar and Mishra, Shailesh and Endres, Christoph and Holz, Thorsten and Fritz, Mario , journal =. Not What You've Signed Up For Compromising Real-World on Pith
  11. [11] Trivedi, Harsh and Khot, Tushar and Hartmann, Mareike and Manku, Ruskin and Dong, Vinty and Li, Edward and Gupta, Shashank and Sabharwal, Ashish and Balasubramanian, Niranjan , booktitle =. 2024 , publisher = on Pith
  12. [12] Mou, Yutao and Xue, Zhangchi and Li, Lijun and Liu, Peiyang and Zhang, Shikun and Ye, Wei and Shao, Jing , booktitle =. 2026 , pages =. doi:10.18653/v1/2026.findings-acl.1850 , url = on Pithdoi:10.18653/v1/2026.findings-acl.1850
  13. [13] Zhao, Haiquan and Yuan, Chenhan and Huang, Fei and Hu, Xiaomeng and Zhang, Yichang and Yang, An and Yu, Bowen and Liu, Dayiheng and Zhou, Jingren and Lin, Junyang and others , journal = on Pith
  14. [14] Zhang, Zhexin and Cui, Shiyao and Lu, Yida and Zhou, Jingzhuo and Yang, Junxiao and Wang, Hongning and Huang, Minlie , journal = on Pith
  15. [15] Building a Foundational Guardrail for General Agentic Systems via Synthetic Data , author =. The Fourteenth International Conference on Learning Representations , year = on Pith
  16. [16] Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and Huang, Chengen and Lv, Chenxu and others , journal = on Pith
  17. [17] Yang, An and Yang, Baosong and Hui, Beichen and Zheng, Bo and Yu, Bowen and Zhou, Chang and Li, Chengpeng and Li, Chengyuan and Liu, Dayiheng and Huang, Fei and others , journal = on Pith
  18. [18] Jing, Huihao and Li, Haoran and Hu, Wenbin and Hu, Qi and Heli, Xu and Chu, Tianshu and Hu, Peizhao and Song, Yangqiu. MCIP : Protecting MCP Safety via Model Contextual Integrity Protocol. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. ... on Pithdoi:10.18653/v1/2025.emnlp-main.62
  19. [19] Wang, Haoyu and Poskitt, Christopher M. and Sun, Jun , year =. Proceedings of the IEEE/ACM 48th International Conference on Software Engineering , eprint = on Pith
  20. [20] Song, Kefan and Qi, Yanjun , booktitle =. 2026 , eprint = on Pith
show 58 more
  1. [21] Han, Seungju and Rao, Kavel and Ettinger, Allyson and Jiang, Liwei and Lin, Bill Yuchen and Lambert, Nathan and Choi, Yejin and Dziri, Nouha , booktitle =. 2024 , url = on Pith
  2. [22] Building a Foundational Guardrail for General Agentic Systems via Synthetic Data , author =. International Conference on Learning Representations , year = on Pith
  3. [23] Xiang, Zhen and Zheng, Linzhi and Li, Yanjie and Hong, Junyuan and Li, Qinbin and Xie, Han and Zhang, Jiawei and Xiong, Zidi and Xie, Chulin and Yang, Carl and Song, Dawn and Li, Bo , booktitle =. 2025 , url = on Pith
  4. [24] Luo, Weidi and Dai, Shenghong and Liu, Xiaogeng and Banerjee, Suman and Sun, Huan and Chen, Muhao and Xiao, Chaowei , booktitle =. 2025 , doi = on Pith
  5. [25] Defeating Prompt Injections by Design , author=. arXiv preprint arXiv:2503.18813 , year= on PithPith's reviewarXiv
  6. [26] Yang, Rui and Fu, Michael and Tantithamthavorn, Kla and Arora, Chetan and Chua, Joey , journal=. 2026 , url= on Pith
  7. [27] Skill Scanner: Security Scanning for. 2026 , howpublished = on Pith
  8. [28] Skill Scanner , author =. 2026 , howpublished = on Pith
  9. [29] 2026 , howpublished = on Pith
  10. [30] Shi, Tianneng and He, Jingxuan and Wang, Zhun and Li, Hongwei and Wu, Linyu and Guo, Wenbo and Song, Dawn , journal = on Pith
  11. [31] Jia, Feiran and Wu, Tong and Qin, Xin and Squicciarini, Anna , booktitle =. The Task Shield: Enforcing Task Alignment to Defend Against Indirect Prompt Injection in. 2025 , doi = on Pith
  12. [32] Wang, Peiran and Li, Ying and Tian, Yuan , journal =. Aligning Provenance with Authorization: A Dual-Graph Defense for on Pith
  13. [33] Kravchenko, Arseny and Liventsev, Vadim and Konstantinov, Innokentii and Iskhakov, Ildar and Kukuy, Matvey , journal = on Pith
  14. [34] Russinovich, Mark , journal =. 2026 , url = on Pith
  15. [35] Le, Minh Nhat and Gondi, Nisarga and Peng, Yibo and Ni, Ronghao and Jia, Limin and Chen, Beidi and Zheng, Haizhong , journal =. Self-Evolving Defense: Continual Security Policy Learning for. 2026 , url = on Pith
  16. [36] Li, Yanjie and He, Xiangyu and Dai, Xuelong and Xiao, Bin , journal =. 2026 , url = on Pith
  17. [37] Zhang, Xiaorui and Cheng, Zhuoran and Liu, Kailin and Sun, Zhaoxi and Fan, Shiyu and Yuan, Tongyu and Yuan, Bin and Qiang, Weizhong and Zou, Deqing , journal = on Pith
  18. [38] Dhouib, Mohamed and Elliker, Clement and Canesse, Alexi and Jenny, Ma. arXiv preprint arXiv:2610.06401 , year = on PithPith's reviewarXiv
  19. [39] Peters-Gill, James and Semler, Avi and Bartsch, Henning and Shumailov, Ilia and Schroeder de Witt, Christian , journal =. Can. 2026 , url = on Pith
  20. [40] Self-Reflection Fine-Tuning: Enhancing Agent Security against Prompt Injection Attacks from Failure Experience , author =. arXiv preprint arXiv:2610.04269 , year = on PithPith's reviewarXiv
  21. [41] Zhang, Yifan and Dai, Yutong and Prabhu, Viraj and Hu, Zhiyuan and Xu, Ran and Chen, Zeyuan , journal =. 2026 , url = on Pith
  22. [42] Blocking at the Boundary: Auditing Long-Horizon Agents against Staged Prompt Injection , author =. arXiv preprint arXiv:2610.05163 , year = on PithPith's reviewarXiv
  23. [43] Readable Before Actionable: Causal Tracing of Indirect Prompt Injection , author =. arXiv preprint arXiv:2610.05295 , year = on PithPith's reviewarXiv
  24. [44] Threat-Preserving Representation Sensitivity in Agent-Security Benchmarks , author =. arXiv preprint arXiv:2610.03585 , year = on PithPith's reviewarXiv
  25. [45] Who Is Your Agent Serving? Provider-Side Indirect Prompt Injection in Proactive Agents , author =. arXiv preprint arXiv:2610.05266 , year = on PithPith's reviewarXiv
  26. [46] Surviving the Router: Optimizing Skill Injections for Retrieval and Execution , author =. arXiv preprint arXiv:2610.08098 , year = on PithPith's reviewarXiv
  27. [47] Towards a Unified Misuse Monitoring Benchmark , author =. arXiv preprint arXiv:2610.07089 , year = on PithPith's reviewarXiv
  28. [48] AdvSim2Real: Training Web Agents Against Adaptive Prompt Injection in a Web World Model , author=. arXiv preprint arXiv:2610.08773 , year= on PithPith's reviewarXiv
  29. [49] FC-SWE: Failure-Conditioned RL for Long-Horizon Software Engineering Agents , author =. arXiv preprint arXiv:2610.07898 , year = on PithPith's reviewarXiv
  30. [50] StateWise: Diagnosing and Repairing Persistent Operational State Before Agent Actions , author =. arXiv preprint arXiv:2610.05241 , year = on PithPith's reviewarXiv
  31. [51] EMHO: EMbodied Agent Harness Optimization via Experience Traces , author =. arXiv preprint arXiv:2610.08432 , year = on PithPith's reviewarXiv
  32. [52] Quantization Effects on Tool-Failure Recovery Vary Across Prompts and Evaluation Designs , author =. arXiv preprint arXiv:2610.07781 , year = on PithPith's reviewarXiv
  33. [53] Wang, Qingzhuo and Wang, CaiYi and Meng, Jinglu and Qin, Ruiyang and Peng, Kunyu and Wei, Zhihua and Shen, Wen , year =. 2609.32428v1 , archivePrefix = on PithPith's reviewarXiv
  34. [54] Zabolotnii, Serhii , year =. Trust Is Not a Score: Runtime Assurance Contracts for High-Risk. 2609.39717v1 , archivePrefix = on PithPith's reviewarXiv
  35. [55] Liu, Zhuo and Li, Moxin and Ma, Zhixin and Shi, Wentao and Wang, Wenjie and Feng, Fuli , journal=. 2026 , url= on Pith
  36. [56] Intrusion Detection for Agentic Processes: Evidence-Based Runtime Monitoring , author=. arXiv preprint arXiv:2610.00151 , year= on PithPith's reviewarXiv
  37. [57] Set Transformer: A Framework for Attention-Based Permutation-Invariant Neural Networks , author =. Proceedings of the 36th International Conference on Machine Learning , pages =. 2019 , url = on Pith
  38. [58] Perceiver: General Perception with Iterative Attention , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , url = on Pith
  39. [59] Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity , author =. Journal of Machine Learning Research , volume =. 2022 , url = on Pith
  40. [60] Zhang, Kaiyuan and Peng, Yuke and Jiang, Ke and Zhang, Yinqian , journal =. 2026 , url = on Pith
  41. [61] Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =. 2023 , url = on Pith
  42. [62] Guo, Wenbo and Zeng, Wei and Liu, Chengwei and Jia, Xiaojun and Xu, Yijia and Tang, Lei and Fang, Yong and Liu, Yang , journal =. 2026 , url = on Pith
  43. [63] Wang, Yue and Liu, Yi and Deng, Gelei and Zhang, Ying and Li, Yuekang and Chen, Zhenyu and Zhang, Leo , journal =. 2026 , url = on Pith
  44. [64] Toolformer: Language Models Can Teach Themselves to Use Tools , author =. Advances in Neural Information Processing Systems 36 , year = on Pith
  45. [65] Ruan, Yangjun and Dong, Honghua and Wang, Andrew and Pitis, Silviu and Zhou, Yongchao and Ba, Jimmy and Dubois, Yann and Maddison, Chris J. and Hashimoto, Tatsunori , booktitle =. Identifying the Risks of. 2024 , url = on Pith
  46. [66] Andriushchenko, Maksym and Souly, Alexandra and Dziemian, Mateusz and Duenas, Derek and Lin, Maxwell and Wang, Justin and Hendrycks, Dan and Zou, Andy and Kolter, Zico and Fredrikson, Matt and Winsor, Eric and Wynne, Jerome and Gal, Yarin and Davies, Xander , booktitle =. 2025 , url = on Pith
  47. [67] Zhang, Hanrong and Huang, Jingyuan and Mei, Kai and Yao, Yifei and Wang, Zhenting and Zhan, Chenlu and Wang, Hongwei and Zhang, Yongfeng , booktitle =. Agent Security Bench (. 2025 , url = on Pith
  48. [68] Inan, Hakan and Upasani, Kartikeya and Chi, Jianfeng and Rungta, Rashi and Iyer, Krithika and Mao, Yuning and Tontchev, Michael and Hu, Qing and Fuller, Brian and Testuggine, Davide and Khabsa, Madian , journal =. Llama Guard:. 2023 , url = on Pith
  49. [69] She, Yining and Liang, Yiliang and Kang, Eunsuk , journal =. Safeguarding. 2026 , url = on Pith
  50. [70] Santos-Grueiro, Igor , journal =. Temporary Authority, Permanent Effects: Commit-Time Authorization for. 2026 , url = on Pith
  51. [71] Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2025 , doi = on Pith
  52. [72] Text Embeddings by Weakly-Supervised Contrastive Pre-training , author =. arXiv preprint arXiv:2212.03533 , year = on PithPith's reviewarXiv
  53. [73] Decoupled Weight Decay Regularization , author =. International Conference on Learning Representations , year = on Pith
  54. [74] Shazeer, Noam , journal = on Pith
  55. [75] Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer , author =. International Conference on Learning Representations , year = on Pith
  56. [76] Gemma 4 Technical Report , author =. arXiv preprint arXiv:2607.02770 , year = on PithPith's reviewarXiv
  57. [77] gpt-oss-120b & gpt-oss-20b Model Card , author =. arXiv preprint arXiv:2508.10925 , year = on PithPith's reviewarXiv
  58. [78] Benchmarking Security Risk Detection and Verification in Open Agentic Skill Ecosystems , author =. arXiv preprint arXiv:2606.00925 , year = on PithPith's reviewarXiv

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment