Cryptography and Security · arXiv:2610.15537 v1 · submitted 11 Oct 2026 · 13 references
Poisoned into Action: Context Attacks on Typed Decision Workflows
The paper says
“Typed decision interfaces restrict models to predefined actions, yet schema-valid choices can still trigger policy-violating state changes. Evaluating context poisoning thus requires connecting model decisions to execution outcomes while accounting for legitimate automation retained by defenses.”
more of the abstract
We introduce an execution-aware protocol tracing poisoned evidence through retrieval, model choice, automated action, and persisted state. We evaluate it with 1,659 fresh calls to Jev-1.13.0 across 20 development tasks in two synthetic workflows, using three repetitions per condition. Under natural retrieval, a standalone payload reaches all 60 trials and produces five security-violating executions, while a carrier-payload pair exposes the payload in 57 trials and produces one violation. An initial source-exclusion gate blocks the pair's observed violation but also withholds ten otherwise correct attacked decisions and sixteen correct clean decisions. Correcting its empty-scope behavior restores twelve clean decisions without additional model calls. Factorial controls show that adding correct evidence reduces wrong executions from 119/120 to 55/120 in the semantic arm and from 60/120 to zero in the neutral arm, suggesting that much of the improvement reflects general evidence value rather than attack-specific protection. Three prespecified contrasts are not statistically supported after Holm correction, so the findings remain exploratory. The protocol provides an auditable framework for studying which poisoning-induced errors reach execution and how much correct automation defenses preserve. Code is available at https://github.com/WuXixiong/jev-poisoning-bench.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 14 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, with conditions. A disciplined execution-aware measurement protocol that cleanly separates exposure, wrong execution, security violation, and lost automation for typed decision workflows; the headline counts rest on private labels over 20 development tasks, so treat the numbers as illustrative of the protocol rather than a robust attack result. This line is wrong
2 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 13 references say.
The paper says, at §4.1, §5.2, Eq. (2)–(4), Appendix E
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.79):
“The public materials therefore do not by themselves support full independent delivery verification.”
Pith says: Objection 1 of 2 · major
The headline empirical claims—payload 5/60 and pair 1/60 security-violating executions, the wrong-execution rows of Tables 4 and 5, and the zero-wrong-execution gate row—are computed in Eq. (2)–(4) from the private allowed-action sets A*_i and security predicates S_i described in §4.1.
more of objection 1
These labels are not part of the public artifact: Appendix E states that the public materials "do not by themselves support full independent delivery verification," and §4.1 acknowledges that the labels were not produced by a blinded multi-reviewer process. Because each task is reference-assisted (the correct meaning of an opaque marker is fixed by the hidden-world assignment rather than determined by the public facts alone), an independent reader cannot recompute the central quantities from the released materials. This is an auditability gap in the single quantity on which the conclusions depend. The fix is concrete and within scope: the labels are generated by fixed synthetic policy rules, so releasing A*_i and S_i per task in the artifact would not compromise the executor's runtime independence; alternatively, a documented blinded re-derivation of the labels from the stated rules would suffice. Without one of these, the 5/60, 1/60, and Table 5 gate counts remain private assertions that no external party can check.
The paper says, at §5.2, Table 4; §4.4
The review did not quote this passage.
Pith says: Objection 2 of 2 · major
The central descriptive contrast (payload 60/60 exposure with 5 violations versus pair 57/60 exposure with 1 violation) is reported as draw-level counts with planned draws as denominators, while §4.4 states that the experimental unit is the 20 tasks. All five payload violations occur on three incident tasks, so the exposure-harm gap the paper emphasizes rests on a handful of task-level …
more of objection 2
… outcomes; under a task-level view the contrast is 3 of 10 incident tasks versus 1 of 10, and a single-task relabeling could change or erase the 5-versus-1 difference. The paper computes conditional task-level bootstrap intervals for its three primary contrasts but not for the descriptive counts that carry the main narrative. A per-task breakdown (which tasks, how many of their three draws were violations) and a simple task-level interval for the H counts would make the sensitivity of the headline result visible and would align the reporting with the paper's own statement of the unit of analysis; the blanket caveat that the findings are exploratory does not substitute for this quantification.
5 minor objections
The paper says, at Abstract, §1, §5.3
Pith says: minor objection 1 of 5
The statement that the source-exclusion gate "blocks the pair's observed violation" rests on a single observed wrong execution; §5.3 qualifies this correctly ("The single ungated pair error is too little evidence for a general ranking"), so the abstract and introduction should carry the same single-event qualifier.
The paper says, at Figure 1
Pith says: minor objection 2 of 5
The embedded string "PoisononJev(An Urgent X-Risk Issue)" appears to be a garbled or leftover placeholder; if not intentional, it should be removed or replaced with a readable example incident title.
show 3 more
The paper says, at Appendix D
Pith says: minor objection 3 of 5
The McNemar result p = 0.1796 is reported without the number of discordant pairs; since the historical reanalysis is presented as a read-only check that readers are expected to verify, reporting the discordant-pair counts would make the computation checkable.
The paper says, at Appendix C, Eq. (8)
Pith says: minor objection 4 of 5
The scope-aware replay rule uses the term "support" ("support is empty") although the surrounding text defines S(R) as the set of registered low-trust principals; the term "support" should be introduced explicitly when S(R) is defined to avoid a forward reference.
The paper says, at §5.2
Pith says: minor objection 5 of 5
The sentence reporting that the six errors in the payload and pair arms have API confidence between 0.19 and 0.29 should state explicitly that these are raw API confidence scores rather than transformed probabilities, consistent with the distinction drawn in §3.1, so that table-only readers do not conflate the two scores.
What would settle it
Run a blinded label audit: two independent policy auditors, seeing only the frozen policy text, the public task worlds, and the correct reference meaning (and not Jev outputs or each other's derivations), produce A⋆i and Si for all 20 tasks. Then rerun the public scoring path (e.g., build_evidence.py --check against the pilot_analysis.json artifacts) with these independently derived labels and compare every W/H/C count in Tables 4 and 5. If the five payload violations, the pair's one violation, and the gate's blocked-execution row reproduce exactly, the label concern is resolved; if any count changes, the harm and protection claims are label-dependent and should be reported as conditional on the authors' private assignment.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 14 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 2 major objections, 5 minor
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 contribution is the measurement protocol, and it is a well-built one. It separates exposure (content visible in context) from wrong execution (action outside the allowed set) from security violation (persisted state breaking a predicate) from lost automation (correct executions a gate withholds). That is a genuinely useful set of distinctions for evaluating context poisoning in typed decision workflows, delivered with real discipline: a mechanical executor that cannot read the answer key, an independent read-only recomputation of all records, frozen plans, hash-checked artifacts, and prespecified contrasts with Holm correction. The CS2 factorial is the sharpest piece; it shows most of the improvement from restoring correct evidence is general evidence value (neutral baseline drops from 60/120 to zero) rather than attack-specific protection. That kills a common attribution error and is worth citing on its own. The empty-scope gate repair (absence of an intervention target is not failed counterfactual support) is a small, concrete insight. The soft spots are real but largely acknowledged. The headline counts, 5/60 violations, the 1/60 pair violation, and the gate blocking it, all depend on the private A*_i and Si labels, which are not in the public bundle and were not produced by a blinded multi-review process. The stress-test concern lands: in these reference-assisted tasks the trusted context does not itself determine the correct action; the hidden-world assignment does. An independent re-derivation of those labels could shift the counts. The authors state this limitation plainly, but it means the central harm/protection numbers are illustrative of the protocol, not settled facts. Second, 20 previously used development tasks from two templates, three draws each, all three primary contrasts null after Holm correction, with headline rates using draws as the denominator while the task is the unit. Acknowledged, and it caps the strength of every empirical claim. Third, one hosted model version, one acquisition window, one frozen contamination pair. Also acknowledged. Note what is not wrong with it: the central claims do not depend on fitted parameters; the calibration temperatures and coverage-matched thresholds are explicitly retrospective diagnostics. No circularity burden there. And the authors are not overclaiming; their claim-scope table and limitations section are more honest than most published work. This is for people building or reviewing robustness evals for typed decision models, especially anyone tempted to stop at flip rates. It deserves a serious referee. I would send it to review, asking for a held-out workflow family, task-level intervals, and an independent blinded label re-derivation in the follow-up.
The strongest objection
The paper's affirmative result—that natural retrieval of a payload yields five security-violating executions, that the pair yields one, and that source-exclusion gating blocks that one while withholding correct executions—is computed through Eq. (2)–(4), where U, H, and C are direct functions of the private allowed-action sets A⋆i and security predicates Si described in §4.1. Those labels are not in the public bundle. The authors explicitly acknowledge that the labels were not produced by a blinded multi-reviewer process (§4.1), and Appendix E states that the public materials do not by themselves support full independent delivery verification. The structure of the synthetic tasks makes this gap load-bearing: each task is a 'reference-assisted decision' in which the correct meaning of an opaque marker is supplied by a corpus document, and in the logged example (Appendix F) even the correct reference entries are low-trust model-visible evidence. The trusted context alone therefore does not determine the correct action; the ground truth is fixed by the authors' hidden-world assignment and policy interpretation. If an independent application of the stated rules to the same public facts assigns even one incident task a different allowed action or a different security-predicate outcome, the counts 5/60, 1/60, and the 0-wrong-execution gate row in Table 5 could change. This is not an accusation; it is an auditability gap in the single quantity on which the conclusion depends.
The reader's weakest assumption is the same one I would name: the central empirical result is measured against private, non-blinded semantic labels. The paper handles this honestly by stating the limitation, and the protocol is otherwise well disclosed: mechanical execution, hash-checked artifacts, prespecified contrasts, and explicit exploratory framing. I do not see an internal inconsistency or a false inference that would justify rejection. The strongest claim is existence-like—these counts were observed—but whether they demonstrate what they claim ('security violations', 'blocked violation') depends on the labels being faithful to the stated policy. Because all three prespecified contrast tests are null after Holm correction, these descriptive counts are the only affirmative support for the conclusion, which raises the stakes of the label audit. If an independent re-derivation reproduces the labels, the paper's scoped claims are credible; if not, Tables 4 and 5 would need re-reporting. That is exactly the conditional posture the reader adopted, so I leave the verdict unchanged.
The referee's report
If the counts hold, the paper makes a useful methodological contribution: it separates exposure, decision error, security violation, and service loss within one auditable record, and it demonstrates that these quantities can diverge sharply (60/60 exposure versus 5/60 harm; wrong executions without violation; a gate that blocks harm while withholding sixteen correct actions). The CS2 factorial logic, which quantifies general evidence value before crediting attack-specific protection, is a substantive analytic contribution that other benchmark efforts should adopt, and the empty-scope repair is a concrete, transferable gate-design insight. The paper's integrity practices are a real strength: a hash-frozen plan, machine-checked artifact lineage, a read-only recomputation of all 1,980 records, prespecified contrasts with Holm correction, and an unusually candid claim-scope table (Appendix B). The central exposure and violation counts do not depend on any fitted parameter; the two temperatures and coverage-matched thresholds are explicitly retrospective diagnostics (CS6, Appendix A).
Pith's reading of the paper (a machine's summary, not the authors')
This paper argues that a schema-valid choice from a typed decision interface can still execute a policy-violating action, so context-poisoning risk must be measured through to persisted state rather than at a changed answer. To make that measurable, the authors introduce an execution-aware protocol with independent correctness and security labels, a mechanical executor that cannot see the answer key, and an auditable record of retrieval, model choice, execution, and state. In a pilot on 20 previously used development tasks in two synthetic workflows, with 1,659 fresh calls to the Jev-1.13.0 decision model, a naturally retrieved payload reached all 60 draws and produced five security-violating executions, while a carrier-payload pair produced one; a source-exclusion gate blocked the pair's violation but withheld 26 correct executions, 12 of which a zero-call scope repair recovered. The paper further shows that restoring correct evidence helps attacked and neutral decisions alike, so much of the observed improvement is general competence rather than attack-specific protection. These findings are explicitly exploratory, with none of the three prespecified contrasts statistically supported after Holm correction; the contribution is the protocol and the exposure-versus-harm-versus-protection distinctions it makes visible.
What carries the argument. The carrying mechanism is the paper's execution-aware protocol, which records four indicator variables per planned draw — $W$ (a wrong available choice), $U$ (a wrong execution), $H$ (a persisted state that violates the security predicate), and $C$ (a correct execution) — scored against private policy labels $A^\star_i$ and a security predicate $S_i$ that the model, retriever, gate, and executor never see. A mechanical executor applies allowed actions to isolated filesystem state without access to the answer key, so an incorrect action is kept distinct from a policy violation, and a gate refusal is scored as lost correct automation rather than as a violation. The delivery study runs on a frozen carrier-payload contamination pair under natural lexical retrieval, and the defense study on a source-exclusion gate $G_0$ that reruns retrieval after removing each registered low-trust principal and demands nonempty, valid, unanimous counterfactual support, together with a scope-aware variant $G_{\mathrm{scope}}$ that permits execution when the registry confirms no low-trust source is represented in the evidence.
Core claim. The central claim is that a changed decision, a wrong executed action, and a security violation are three different events, and that context-poisoning risk in typed decision workflows must be read as a chain: whether poisoned text is present in the corpus, whether retrieval exposes it, whether the model chooses a wrong action, whether that action executes, what persisted state it leaves behind, and how much correct automation a defense preserves while blocking harm. Instantiating this chain on two synthetic workflows with fixed facts, policy, and action definitions, the paper observes under natural retrieval that a standalone payload reached all 60 draws and produced five security-violating executions across three incident tasks, while a carrier-payload pair exposed the payload in 57 of 60 draws and produced one such violation. A source-exclusion gate removed the pair's single observed violation but withheld 26 otherwise correct executions; a scope-aware repair that treats registry-confirmed absence of a low-trust source as valid support rather than failure restored 12 of the 16 withheld clean decisions without any new model calls. A factorial control showed that restoring the correct reference cut wrong executions in the semantic arm from 119/120 to 55/120 and in the neutral arm from 60/120 to zero, indicating that most of the improvement is general evidence value rather than attack-specific protection.
The premise it rests on. Everything measured depends on the authors' offline policy rules being the correct ground truth for correctness and security in two synthetic workflows; if those rules do not capture what a real deployment's policies would count as damage, the observed exposure, harm, and protection gaps will not transfer.
If the paper is right (Pith's inferences, not the authors')
A decision-flip rate alone misstates poisoning risk: the same semantic content produced 10 wrong executions through state text, zero through appended documents, and 60 through outer instructions, so input authority is part of the threat model, not a nuisance variable.
Defenses should be evaluated by counting both prevented violations and withheld correct automation: the source gate eliminated the pair's only observed violation while dropping correct clean executions from 60/60 to 44/60, and the empty-scope repair recovered 12 of those at no model cost.
Restoring correct evidence improves both attacked and neutral decisions, so benchmarks that credit all such gains to robustness will systematically overstate attack-specific protection.
Corpus presence and request exposure are not harm: payload text was visible in 57 to 60 of 60 draws, yet only six executions across all natural-retrieval conditions left a policy-violating persisted state.
Beyond the paper: if the exposure-versus-harm gap holds beyond these two templates, the natural next step for typed-decision benchmarks is to report every poisoning result as an execution chain — presence, exposure, wrong execution, violated state, withheld correct automation — and this protocol is the template for doing so.
Beyond the paper: the gate's empty-scope bug is likely one instance of a general defense failure in which 'no intervention target found' is treated as failed counterfactual support; other provenance-based filters should be audited with the same zero-call registry replay.
Beyond the paper: the evidence-competition result suggests a testable prediction — on workflows where retrieval already surfaces correct references, selective-execution defenses will show little marginal protection, implying that retrieval quality and gating should be tuned as separate levers.
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
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.
The paper
Abstract
Typed decision interfaces restrict models to predefined actions, yet schema-valid choices can still trigger policy-violating state changes. Evaluating context poisoning thus requires connecting model decisions to execution outcomes while accounting for legitimate automation retained by defenses. We introduce an execution-aware protocol tracing poisoned evidence through retrieval, model choice, automated action, and persisted state. We evaluate it with 1,659 fresh calls to Jev-1.13.0 across 20 development tasks in two synthetic workflows, using three repetitions per condition. Under natural retrieval, a standalone payload reaches all 60 trials and produces five security-violating executions, while a carrier-payload pair exposes the payload in 57 trials and produces one violation. An initial source-exclusion gate blocks the pair's observed violation but also withholds ten otherwise correct attacked decisions and sixteen correct clean decisions. Correcting its empty-scope behavior restores twelve clean decisions without additional model calls. Factorial controls show that adding correct evidence reduces wrong executions from 119/120 to 55/120 in the semantic arm and from 60/120 to zero in the neutral arm, suggesting that much of the improvement reflects general evidence value rather than attack-specific protection. Three prespecified contrasts are not statistically supported after Holm correction, so the findings remain exploratory. The protocol provides an auditable framework for studying which poisoning-induced errors reach execution and how much correct automation defenses preserve. Code is available at https://github.com/WuXixiong/jev-poisoning-bench.
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 · 2026-10-14
The authors' main claim, as the classifier picked it out 0.36
“Typed decision interfaces restrict models to predefined actions, yet schema-valid choices can still trigger policy-violating state changes.”
How the main claim is stated: asserted 0.93
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.97 · human participants 0.04 · controlled comparison 0.95 · randomized assignment 0.04 · states quantitative result 0.99 · compares with prior results 0.74 · mentions code or data availability 0.98
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.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Typed decision interfaces restrict models to predefined actions, yet schema-valid choices can still trigger policy-violating state changes. | background 0.83 | 0.10 | 0.22 | asserted 0.93 |
| Evaluating context poisoning thus requires connecting model decisions to execution outcomes while accounting for legitimate automation retained by defenses. | gap 0.64 | 0.13 | 0.07 | asserted 0.96 |
| We introduce an execution-aware protocol tracing poisoned evidence through retrieval, model choice, automated action, and persisted state. | contribution 0.84 | 0.94 | 0.07 | asserted 0.99 |
| We evaluate it with 1,659 fresh calls to Jev-1.13.0 across 20 development tasks in two synthetic workflows, using three repetitions per condition. | method 0.99 | 0.96 | 0.03 | asserted 0.99 |
| Under natural retrieval, a standalone payload reaches all 60 trials and produces five security-violating executions, while a carrier-payload pair exposes the payload in 57 trials and produces one violation. | result 1.00 | 0.95 | 0.11 | asserted 1.00 |
| An initial source-exclusion gate blocks the pair's observed violation but also withholds ten otherwise correct attacked decisions and sixteen correct clean decisions. | result 1.00 | 0.93 | 0.07 | asserted 1.00 |
| Correcting its empty-scope behavior restores twelve clean decisions without additional model calls. | result 0.94 | 0.92 | 0.06 | asserted 1.00 |
| Factorial controls show that adding correct evidence reduces wrong executions from 119/120 to 55/120 in the semantic arm and from 60/120 to zero in the neutral arm, suggesting that much of the improvement reflects general evidence value rather than attack-specific protection. | interpretation 0.74 | 0.95 | 0.20 | hedged possibility 0.93 |
| Three prespecified contrasts are not statistically supported after Holm correction, so the findings remain exploratory. | limitation 0.98 | 0.93 | 0.08 | asserted 0.99 |
| The protocol provides an auditable framework for studying which poisoning-induced errors reach execution and how much correct automation defenses preserve. | contribution 0.87 | 0.81 | 0.36 | asserted 0.99 |
| Code is available at https://github.com/WuXixiong/jev-poisoning-bench. | contribution 0.86 | 0.54 | 0.02 | asserted 0.96 |
What kind of paper the classifier reports this is: method or tool 0.68 · evidence from observation or experiment 0.99
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, 2026-10-14: 1 request, 8,222 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 (3)
Figures are the authors', reproduced from the paper.
References (13)
- [1] Edoardo Debenedetti, Jie Zhang, Mislav Balunovic, Luca Beurer-Kellner, Marc Fischer, and Florian Tramèr. Agentdojo: A dynamic environment to evaluate prompt injection attacks and defenses for llm agents.Advances in neural information processing systems, 37:82895–82920, 2024 on Pith
- [2] Yuyang Gong, Zhuo Chen, Jiawei Liu, Miaokun Chen, Fengchang Yu, Wei Lu, XiaoFeng Wang, and Xiaozhong Liu. Topic-FlipRAG: Topic-orientated adversarial opin- ion manipulation attacks to retrieval-augmented gener- ation models. In34th USENIX Security Symposium, on Pith
- [3] Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. Not what you’ve signed up for: Compromising real- world LLM-integrated applications with indirect prompt injection. arXiv:2302.12173v2, 2023. URL https: //arxiv.org/abs/2302.12173v2 on PithPith's reviewarXiv
- [4] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Wein- berger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 ofProceedings of Ma- chine Learning Research, pages 1321–1330, 2017. URL https://proceeding... on Pith
- [5] Jianyi Hu, Hangtao Zhang, Yi Liu, Yeqi Zeng, Li Zeng, Xianlong Wang, Rui Wang, and Leo Yu Zhang. Je- vAdvBench: A benchmark and black-box attacks for reinforcement learning for calibrated decisions models. arXiv:2609.31142v1, 2026. URL https://arxiv.org/ abs/2609.31142v1 on PithPith's reviewarXiv
- [6] Yu Sun, Junhao Xu, Jiajia Shi, and Zijin Yang. Type- safe is not error-free: A constrained decision head follows the option name, not the rubric bound to it. arXiv:2609.26758v2, 2026. URL https://arxiv.org/ abs/2609.26758v2 on PithPith's reviewarXiv
- [7] Tiantong Wu and Wei Yang Bryan Lim. Decision hi- jacking: Prompt injection attacks on Jev’s typed prob- abilistic decisions. arXiv:2609.28613v1, 2026. URL https://arxiv.org/abs/2609.28613v1 on PithPith's reviewarXiv
- [8] Chong Xiang, Tong Wu, Zexuan Zhong, David Wagner, Danqi Chen, and Prateek Mittal. Certifiably robust rag against retrieval corruption. In2026 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 392–410, 2026 on Pith
- [9] Zixiang Xu. JevOut: Natural context can flip decision models. arXiv:2609.30243v1, 2026. URLhttps://arxi v.org/abs/2609.30243v1 on PithPith's reviewarXiv
- [10] Qiusi Zhan, Zhixiang Liang, Zifan Ying, and Daniel Kang. Injecagent: Benchmarking indirect prompt injec- tions in tool-integrated large language model agents. In Findings of the Association for Computational Linguis- tics: ACL 2024, pages 10471–10506, 2024 on Pith
- [11] Huichi Zhou, Kin-Hei Lee, Zhonghao Zhan, Yue Chen, Zhenhao Li, Zhaoyang Wang, Hamed Haddadi, and Emine Yilmaz. TrustRAG: Enhancing robustness and trustworthiness in retrieval-augmented generation. arXiv:2501.00879v3, 2025. URL https://arxiv.org/ abs/2501.00879v3 on PithPith's reviewarXiv
- [12] Wei Zou, Runpeng Geng, Binghui Wang, and Jinyuan Jia. {PoisonedRAG}: Knowledge corruption attacks to{Retrieval-Augmented} generation of large language models. In34th USENIX Security Symposium (USENIX Security 25), pages 3827–3844, 2025. 9 A Additional diagnostics A.1 Calibrati... on Pith
- [2025] URL https://www.usenix.org/conference/us enixsecurity25/presentation/gong-yuyang on Pith
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.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment