Computation and Language · arXiv:2611.02522 v1 · submitted 2 Nov 2026 · 34 references
Ranked but Not Stopped: Detecting Prompt-Injection Hijacks from Typed Decision Models' Output Probabilities
The paper says
“Programs that use a language model as a classifier act on its most probable label. Text injected into the input can move that label to one the attacker chooses, a hijack.”
more of the abstract
A monitor that does not know the attacker's label can read the model's own confidence and raise an alarm when the top probability is low. We test this monitor on AG News topic classification with five systems, three of them served through a provider's API. Pooled over five families of injected text, low confidence ranks hijacked answers above clean ones on four of the five systems (AUROC 0.86-0.90). Ranking them does not stop many of them. A threshold set so that at most 5% of clean answers to separate calibration articles alarm falls among the least confident clean answers, and every more confident hijack passes. In our main gate run on the three API-served systems, the monitor misses 51-70% of the hijacks caused by a plain injected note (per-system point estimates). When the note adds one handwritten sentence that calls the target certain and asks for full probability, the monitor misses 99-100% of the hijacks, although low confidence still ranks them at AUROC 0.70-0.79. Confidence monitors should be tested at their operating alarm rate, against attackers who claim certainty.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 3 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
The passes disagreed. Reader: holds up. Skeptic: agrees with the reader. Referee: major revision.
A solid empirical caution that AUROC overstates what a confidence gate catches at a 5% clean-alarm rate, with a headline plain-note miss range that is less precise than the abstract implies. This line is wrong
1 major objection · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 34 references say.
The paper says, at Abstract; §3.5 and Table 23; §7
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.62):
“Such collapsed ranges and the intervals at the 1% target are bootstrap sensitivity ranges without established coverage.”
Pith says: Objection 1 of 1 · major
The headline plain-note missed share, 51–70% (abstract, §3.5/Table 4, contribution 2, §7), is a single-calibration-draw point estimate whose joint 95% sensitivity ranges (Table 23) span 21–67% (Kimi-K3), 48–100% (Qwen3.8-Flash), and 34–95% (gpt-6-luna); at the lower ends on two systems (21%, 34%) the gate would stop three-quarters to two-thirds of plain-note hijacks, which would materially weaken the title-level "ranked but not stopped" claim for this attack.
more of objection 1
The measurement is fragile by construction: the threshold is the 96th smallest of 100 clean calibration scores (Section 2), so its location is determined by roughly the five least confident calibration calls, and the paper's own Table 23 and Table 20 notes say collapsed intervals are bootstrap sensitivity ranges without established coverage. The corroborating "42–76% across four runs" (Table 5, §7) is not four independent replications: the calibrated and fresh-search runs share the same 100 calibration articles (and, on Kimi-K3 and Qwen3.8-Flash, the same evaluation articles and note position), and the demand and search runs set their thresholds on their own clean calls. The conclusion states "it then misses 51–70%" without the point-estimate qualifier that the abstract carries, and "misses most" is not supported at the paper's own confidence level for the plain note. The certain-note part of this concern does not land: its missed share (99–100%) has intervals [95,100]/[100,100]/[98,100] and replicates under own-clean-call thresholds, 1%/10% alarm targets (Table 20), and the top-five bound (Table 15). The fix is within scope: report the missed shares with their interval span at every appearance in abstract, contributions, and conclusion; temper "most" to "a large minority to nearly all"; state the shared resources among the four runs; or re-pin the estimate with a larger calibration pool.
4 minor objections
The paper says, at Figure 2
Pith says: minor objection 1 of 4
The histograms give no per-group sample sizes and no bin alignment; since the threshold line (top probability 0.62–0.70, Table 23) falls inside a bin, the share of hijacks above the threshold cannot be read from the figure, and the visual claim that "most hijacks pass" is precisely the quantity Table 23 leaves uncertain. Add panel-specific group sizes and state the bin edges.
The paper says, at §4.2
Pith says: minor objection 2 of 4
The sentence "criteria (ii) and (iv) hold; (i) and (iii) do not, with 3 systems against 4 and 2 against 4" forces the reader to guess which count belongs to which criterion; writing "(i) holds on 3 of 5 systems; (iii) on 2 of 5" would remove the ambiguity.
show 2 more
The paper says, at Tables 20 and 23
Pith says: minor objection 3 of 4
The headers call the bracketed quantities "95% intervals" while the notes say collapsed ranges are bootstrap sensitivity ranges without established coverage; renaming the headers to "95% sensitivity range" would keep readers from treating [100,100] as a population bound.
The paper says, at §3.5 and Table 20
Pith says: minor objection 4 of 4
Given the paper's own prescription to evaluate at the deployment alarm rate, reporting the plain-note missed share at additional alarm targets (1% and 10%), as Table 20 does for the searched text, would show whether the plain-note result is as alarm-rate-sensitive as the searched-text result; at present the alarm-rate dependence of the plain note is not shown.
What would settle it
Re-derive the gate from a larger calibration pool: pool the clean low-confidence scores of the 100 calibration and 200 evaluation articles (n=300), set the threshold at k = 286 for a 5% target, and recompute the plain-note and certain-note missed shares on the 200 evaluation articles for gpt-6-luna, Kimi-K3 and Qwen3.8-Flash, with 95% intervals from joint resamples of the 300-article pool. If the plain-note interval lower bound stays above roughly 35% on all three models, the 51-70% range reflects the clean-call geometry rather than one calibration draw; if it falls below roughly 30% on any model, the headline range should be reported as a materially uncertain point estimate, while the certain-note share (expected to remain at or above 98%) and the operating-point recommendation would be unaffected. The ancillary files or a re-run of the public pipeline provide the per-call scores.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 3 Nov 2026
- Reader holds up
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 1 major objection, 4 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
Jianan — this one is worth your time. The paper measures a simple confidence gate (alarm when the top probability is low) on five typed decision systems, three of them API-served, and shows that ranking performance does not translate into catching hijacks at a 5% clean-alarm operating point: the gate misses 51–70% of plain-note hijacks and 99–100% of hijacks from a one-sentence "certain note" on the API-served models. The new piece is the operating-point measurement itself, not the general warning — Carlini et al. already made the low-FPR point — but the concrete numbers on these systems, the gate calibrated on separate clean articles, and the certain-note breakdown add something real.
The paper does several things well. The methods are specific enough to reimplement, the appendices are thorough, and the limitations section is unusually honest: it labels the miss rates as point estimates, notes the intervals cover article sampling only, and flags the unversioned API models and the single prompt per system. The certain-note result is robust — the 99–100% misses replicate across runs and survive the top-five reconstruction bound.
The soft spot is the headline plain-note range. With n=100 calibration articles, the 5% threshold is the 96th smallest clean score, so the plain-note miss-rate measurement is pinned by the five least confident clean calls. Table 23's joint intervals span 21–100% per API model, and the abstract's 51–70% is a point estimate that could shift with a larger calibration pool. The four-run corroboration (42–76%) is weaker than it looks: the calibrated and fresh-search runs share the same 100 calibration articles on Kimi-K3 and Qwen3.8-Flash, and the demand and search runs set their thresholds on their own clean calls. The authors do say "point estimates," but the abstract still leads with the range, and a careful reader should not treat it as a stable measurement. I don't think this sinks the paper: the qualitative claim — AUROC ranking does not imply a low miss rate at 5% clean alarms — holds even if true values sit near the interval lower ends, and the certain-note result is not in question.
Who is this for? Anyone building confidence-based guardrails on typed decision models, and anyone evaluating detectors with AUROC alone. It deserves a serious referee: the measurement design is reproducible and the caution is specific enough to act on. I would accept it for peer review, with the revision request being to shrink the calibration pool or present the plain-note range with intervals front and center rather than as a bare point-estimate range in the abstract.
The strongest objection
The gate's operating point is a single small order statistic: with n=100 calibration articles the 5-percent threshold is the 96th smallest clean score (Section 2), so the entire miss-rate measurement is pinned by roughly five clean calibration calls. Table 23's joint intervals on the plain-note missed share span 21-67% (Kimi-K3), 48-100% (Qwen3.8-Flash) and 34-95% (gpt-6-luna); at the study's own confidence level, missing a small minority and missing nearly all are both inside the interval for every API model. The abstract's 51-70% is thus not a statistically pinned range, and the four-run corroboration (42-76%, Table 5) is weaker than it reads: the calibrated and fresh-search runs share the same 100 calibration articles on Kimi-K3 and Qwen3.8-Flash, and the demand and search runs use own-clean-call thresholds, so the range is not four independent replications of one measurement. The authors do label these as point estimates (Table 23 note; Section 8), so this is a precision concern, not a hidden error. The certain-note result is not in question: intervals collapse to [100,100]/[98,100], it replicates at 1% and 10% alarm targets for the searched text (Table 20), and it survives the top-five reconstruction bound (Table 15). What remains unresolved is whether the specific 51-70% range, and the headline reading that the gate misses most plain-note hijacks, is stable under a larger calibration pool; if true values sat near the interval lower ends, the quantitative claim would overstate the miss rate even though the qualitative claim (AUROC 0.80-0.91 ranking does not imply a low miss rate at 5% clean alarms) would still stand.
The central claim is two-part: (1) on the three API-served models, a 5%-alarm confidence gate misses many plain-note and nearly all certain-note hijacks even though low confidence ranks hijacks above clean calls; (2) monitors should therefore be evaluated at their operating alarm rate, against certainty-claiming attacks. Part (2) is a methodological recommendation that holds regardless of the empirical numbers, so it is not at risk. Part (1)'s certain-note half is strongly supported: collapsed bootstrap intervals, replication across four runs and at 1%/10% alarm targets for the searched text, robustness to the top-five bound, and a second-phrasing/second-model check that the hijacks themselves are real. The plain-note half is where the evidence is softest: Table 23 intervals span 21-100% per system, and the 51-70% headline is a point estimate from one 100-article calibration draw. I checked the internal arithmetic (hijack times missed equals unflagged in Table 4; cross-run ranges in Table 5 are consistent with Tables 19, 23 and 25; the AUROC-gate geometry is coherent given a clean-pmax distribution with a low-confidence tail below 0.96 that sets the 0.62-0.70 threshold) and found no inconsistency. The reader's weakest assumption (deployment exchangeability) is real but largely internalized: the paper scopes to AG News and its advice is to re-calibrate at the deployment's clean alarm rate, so a shifted deployment would invalidate the specific numbers, not the argument. My sharper concern is the in-study precision of the quantitative headline, which is adjacent to the reader's note about wide bootstrap intervals but is about calibration-draw sensitivity rather than transfer. It does not change the verdict: ACCEPT remains right, with the qualitative claim and the recommendation intact, and a single larger-pool calibration check would settle whether the 51-70% range itself is stable.
The referee's report
If the findings hold, the paper makes a useful, partly counterintuitive point: a ranking AUROC over hijacked versus clean calls can be high while a gate at a 5% clean-alarm rate misses most hijacks, because the gate is set by the least confident clean answers and a monitor-aware attacker can push hijack confidence above that threshold. The best-supported result is the certain note: the missed share reaches 99–100% in the calibrated run, with intervals [95,100]/[100,100]/[98,100] that survive joint calibration-and-evaluation resampling, own-clean-call thresholds (92–96%), 1% and 10% alarm targets (Table 20), and the top-five reconstruction bound (Table 15). The ranking-versus-gate divergence also holds at the lower ends of the plain-note intervals, so the qualitative claim does not depend on the headline point estimates. The reporting standards are a strength: the conformal index and its guarantee are stated correctly (Section 2), outside-plan analyses are listed explicitly (Appendix L), reconstruction assumptions are exposed and bounded (Section 3.7), and ancillary files are promised for every printed number.
Pith's reading of the paper (a machine's summary, not the authors')
This paper asks whether a program that uses a language model as a classifier can police itself against prompt-injection hijacks by watching the probabilities that come back with every answer, at no extra cost. The authors find that low confidence, one minus the probability of the top label ($1-p_{\max}$), ranks hijacked answers above clean ones on four of the five systems tested on AG News (pooled AUROC 0.86–0.90). Ranking is not stopping: when the same signal runs as a gate whose threshold lets at most 5% of clean answers alarm, the gate misses 51–70% of hijacks from a plain injected note and 99–100% of hijacks from a note that adds one handwritten sentence declaring the target certain, on the three API-served models. The two jobs part ways because clean answers are mostly near certain, so a 5% alarm threshold sits among the few least-confident clean answers while slightly less confident hijacks rank above them. The paper's practical conclusion is that confidence monitors should be tested at the alarm rate they will run at, not judged by ranking AUROC alone, and against attackers who claim certainty.
What carries the argument. The central objects are the typed decision model, an LLM asked to answer with one of $K$ labels whose output probabilities the calling program gets with every call, and the low-confidence trace $1-p_{\max}$, one minus the probability of the chosen label, the standard baseline for flagging misclassified inputs. The argument is carried by the gate rule: alarm when low confidence exceeds the $k$th smallest of 100 clean calibration scores with $k=\lceil 0.95(n-1)\rceil+1$, the split-conformal index that keeps at most 5% of exchangeable clean calls alarming. The paper's key move is to measure the trace two ways that are usually conflated, the AUROC that ranks hijacks against clean calls and the missed share, the fraction of hijacks that pass at the 5% operating point, and to show these come apart because clean answers cluster near certain on the API-served models. What decides whether a hijack passes is not how it ranks among clean calls but where the least-confident clean answers sit relative to the threshold, and where the injected text pushes the hijack's top probability; a handwritten sentence that claims the target is certain pushes it above the threshold.
Core claim. On its own terms, the discovery is a divergence between two jobs of a confidence score on a typed decision model, a program that acts on the model's most probable label and receives the label probabilities with every call. The first job, ranking, works: low confidence places hijacked calls above the clean calls of the same articles on four of the five systems, with pooled AUROC 0.86–0.90. The second job, stopping, fails: a gate that alarms when low confidence passes a threshold calibrated on 100 separate clean articles, so that at most 5% of clean calls alarm, misses 51–70% of a plain injected note's hijacks and 99–100% of a certain note's hijacks on gpt-6-luna, Kimi-K3 and Qwen3.8-Flash, though these are per-system point estimates with wide intervals. Low confidence still ranks the certain-note hijacks above clean answers, at AUROC 0.70–0.79, so the ranking and the gate disagree about the same calls. The threshold lands at top probability 0.62–0.70, among the few least-confident clean answers, because most clean answers sit near certain, and a hijack with top probability around 0.80 is more confident than the threshold and passes; a second model and a second phrasing of the question also return the attacker's target on 99–100% of certain-note hijacks, so those cross-checks do not rescue the gate.
The premise it rests on. The missed-share figures transfer to a deployment only if the deployment's clean-call distribution is exchangeable with the 100 calibration and 200 evaluation AG News articles under the same prompt and model snapshot; the paper itself states that its intervals cover article sampling only, that the API-served models have no dated snapshot, and that one prompt per system understates sensitivity to wording.
If the paper is right (Pith's inferences, not the authors')
A confidence monitor's AUROC is not a deployment certificate: the paper reports AUROC 0.80–0.91 for plain-note hijacks while the same signal's gate at 5% clean alarms misses 51–70% of them, so monitors should be reported at their operating alarm rate.
Attackers do not need gradient optimization: one handwritten sentence that claims the target is certain and asks for full probability defeats the calibrated gate on 99–100% of certain-note hijacks across the three API-served models, and text found by a simple coordinate search does no better.
Cross-checks do not rescue the gate: a second model and a second phrasing of the question both return the attacker's target on 99–100% of certain-note hijacks, and a disagreement trace between the Choice and a Boolean never beats low confidence on AG News.
The gate's failure is a geometric fact about the operating point: a threshold set to a low clean-alarm rate is pinned by the least-confident clean answers, so where the clean distribution sits determines whether a slightly less confident hijack is ranked yet still allowed through.
Traces chosen on one task transfer unevenly: the fixed trace stays above chance on all five systems on the easy DBpedia-14 subset but on only three of five on the harder Yahoo Answers subset, so deployers should check the trace against labeled hijacks from their own task.
The same ranked-but-not-stopped geometry should apply to any guardrail that thresholds its own confidence on untrusted input — LLM judges, routers and agent loops that act on their own probability — since those systems also read their score from the model under attack; the paper raises this possibility in its discussion but does not test it.
Appendix K's within-article result, that the returned target probability rises with the probability the attacker requests, suggests a cheap detector the paper measures but does not propose: flag texts that demand a named label at near-1.0 probability, and compare requested against returned values.
A deployment whose clean answers are less confident than AG News's, from harder tasks or hedge-inducing prompts, should see the gate's threshold move up and, by the paper's own geometry, catch more hijacks; the open-bosun results show the divergence can also invert, with hijacks about as confident as clean answers and AUROC near chance.
Because the calibrated run's breakdown shows that merely claiming certainty, even without repeating the target or demanding full probability, raises the missed share on every system, input-side defenses could target certainty-claiming phrasing as a specific pattern rather than waiting for optimized adversarial text.
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
Programs that use a language model as a classifier act on its most probable label. Text injected into the input can move that label to one the attacker chooses, a hijack. A monitor that does not know the attacker's label can read the model's own confidence and raise an alarm when the top probability is low. We test this monitor on AG News topic classification with five systems, three of them served through a provider's API. Pooled over five families of injected text, low confidence ranks hijacked answers above clean ones on four of the five systems (AUROC 0.86-0.90). Ranking them does not stop many of them. A threshold set so that at most 5% of clean answers to separate calibration articles alarm falls among the least confident clean answers, and every more confident hijack passes. In our main gate run on the three API-served systems, the monitor misses 51-70% of the hijacks caused by a plain injected note (per-system point estimates). When the note adds one handwritten sentence that calls the target certain and asks for full probability, the monitor misses 99-100% of the hijacks, although low confidence still ranks them at AUROC 0.70-0.79. Confidence monitors should be tested at their operating alarm rate, against attackers who claim certainty.
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 · 2026-11-03
The authors' main claim, as the classifier picked it out 0.74
“Ranking them does not stop many of them.”
How the main claim is stated: asserted 0.99
What the abstract itself states, as probabilities: claims first 0.03 · states limitation 0.96 · human participants 0.06 · controlled comparison 0.73 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.35 · 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.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Programs that use a language model as a classifier act on its most probable label. | background 1.00 | 0.05 | 0.02 | asserted 1.00 |
| Text injected into the input can move that label to one the attacker chooses, a hijack. | background 1.00 | 0.07 | 0.04 | asserted 0.95 |
| A monitor that does not know the attacker's label can read the model's own confidence and raise an alarm when the top probability is low. | method 0.80 | 0.15 | 0.09 | asserted 0.84 |
| We test this monitor on AG News topic classification with five systems, three of them served through a provider's API. | method 0.93 | 0.94 | 0.03 | asserted 0.98 |
| Pooled over five families of injected text, low confidence ranks hijacked answers above clean ones on four of the five systems (AUROC 0.86-0.90). | result 1.00 | 0.96 | 0.15 | asserted 1.00 |
| Ranking them does not stop many of them. | result 0.76 | 0.84 | 0.34 | asserted 0.99 |
| A threshold set so that at most 5% of clean answers to separate calibration articles alarm falls among the least confident clean answers, and every more confident hijack passes. | result 0.68 | 0.86 | 0.14 | asserted 0.99 |
| In our main gate run on the three API-served systems, the monitor misses 51-70% of the hijacks caused by a plain injected note (per-system point estimates). | result 1.00 | 0.96 | 0.23 | asserted 1.00 |
| When the note adds one handwritten sentence that calls the target certain and asks for full probability, the monitor misses 99-100% of the hijacks, although low confidence still ranks them at AUROC 0.70-0.79. | result 1.00 | 0.95 | 0.16 | asserted 1.00 |
| Confidence monitors should be tested at their operating alarm rate, against attackers who claim certainty. | future work 0.63 | 0.11 | 0.63 | asserted 0.90 |
What kind of paper the classifier reports this is: empirical study 0.96 · 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, 2026-11-03: 1 request, 7,533 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 (2)
References (34)
- [3] Md Anas Biswas. Confidently wrong: Severity-aware calibration of prompt-injection detectors under attack shift.arXiv preprint arXiv:2606.22659, on PithPith's reviewarXiv
- [7] Sarthak Choudhary, Divyam Anshumaan, Nils Palumbo, and Somesh Jha. How not to detect prompt injections with an LLM.arXiv preprint arXiv:2507.05630, on PitharXiv
- [8] Reuben Feinman, Ryan R. Curtin, Saurabh Shintre, and Andrew B. Gardner. Detecting adversarial samples from artifacts.arXiv preprint arXiv:1703.00410, on PithPith's reviewarXiv
- [9] 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 preprint arXiv:2302.12173, on PithPith's reviewarXiv
- [10] 13 Ruoqi Guo, Yi Liu, Gelei Deng, Yuekang Li, Lida Zhao, Yutao Wu, Simin Chen, Ying Zhang, and Leo Yu Zhang. Just ask Jev: Reinforcement learning for calibrated decisions as a zero-shot detector of AI alignment failures.arXiv preprint arXiv:2609.29429, on PithPith's reviewarXiv
- [11] Sarthak Gupta and Priyanka Tembey. Are you reading the right thing? Rethinking internal signals for prompt-injection detection.arXiv preprint arXiv:2610.19452, on PithPith's reviewarXiv
- [12] Hugging Face model card. Accessed 2026-10-12. Pengcheng He, Jianfeng Gao, and Weizhu Chen. DeBERTaV3: Improving DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing. InThe Eleventh International Conference on Learning Representations, on Pith
- [13] Blog post. Jianyi Hu, Hangtao Zhang, Yi Liu, Yeqi Zeng, Li Zeng, Xianlong Wang, Rui Wang, and Leo Yu Zhang. JevAdvBench: A benchmark and black-box attacks for reinforcement learning for calibrated decisions models.arXiv preprint arXiv:2609.31142v1, 2026a. Jianyi Hu, Hangtao Zh... on PithPith's reviewarXiv
- [15] Qianlong Lan, Anuj Kaul, Vinothini Pandurangan, and Indranil Sanyal. From detection to decision: Calibration and deferral for prompt-injection detection.arXiv preprint arXiv:2610.20907, on PithPith's reviewarXiv
- [16] Moritz Laurer, Wouter van Atteveldt, Andreu Casas, and Kasper Welbers. Building efficient universal classifiers with natural language inference.arXiv preprint arXiv:2312.17543, on PithPith's reviewarXiv
- [17] Yubo Li, Yidi Miao, Ramayya Krishnan, and Rema Padman. JEV-as-a-judge: Accept when confident, escalate when unsure.arXiv preprint arXiv:2609.26550, on PithPith's reviewarXiv
- [18] Lijia Liu, Takumi Kondo, Kyohei Atarashi, Koh Takeuchi, Jiyi Li, Shigeru Saito, and Hisashi Kashima. Counterfactual evaluation for blind attack detection in LLM-based evaluation systems.arXiv preprint arXiv:2507.23453, on PitharXiv
- [19] Yixuan Liu. Evaluating System One models for agent security decisions: Reliability, calibration, and selective automation.arXiv preprint arXiv:2609.33401, on PithPith's reviewarXiv
- [20] Hugging Face model card. Accessed 2026-10-12. Milad Nasr, Nicholas Carlini, Chawin Sitawarin, Sander V. Schulhoff, Jamie Hayes, et al. The attacker moves second: Stronger adaptive attacks bypass defenses against LLM jailbreaks and prompt injections.arXiv preprint arXiv:2510.09023, on PithPith's reviewarXiv
- [21] Shiyu Ni, Keping Bi, Lulu Yu, and Jiafeng Guo. Are large language models more honest in their probabilistic or verbalized confidence?arXiv preprint arXiv:2408.09773, on PithPith's reviewarXiv
- [22] Public beta; modelgpt-6-luna. Accessed 2026-10-10. Riccardo Porcedda. Jev thinks “I don’t know”, but doesn’t say it: Introducing Sys1Cal-v1 dataset for probability calibration.arXiv preprint arXiv:2609.35342, on PithPith's reviewarXiv
- [23] Franziska Rubenbauer, Sebastian Steindl, Patrick Levi, Daniel Loebenberger, and Ulrich Schäfer. Detection of adversarial prompts with model predictive entropy. InFindings of the Association for Computational Linguistics: EACL 2026, pp. 1979–1993. Association for Computational ... on Pith
- [24] doi: 10.18653/v1/2026.findings-eacl.103. URLhttps://aclanthology.org/2026.findings-eacl.103/. Lijuan Tang and Yuemeng Zheng. Typed decision models: An early evidence audit and evaluation checklist. arXiv preprint arXiv:2609.32160, on PithPith's reviewarXiv
- [25] Mikhail Terekhov, Alexander Panfilov, Daniil Dzenhaliou, Caglar Gulcehre, Maksym Andriushchenko, Ameya Prabhu, and Jonas Geiping. Adaptive attacks on trusted monitors subvert AI control protocols. arXiv preprint arXiv:2510.09462, on PitharXiv
- [26] Mengxiao Wang, Yuxuan Zhang, and Guofei Gu. PromptSleuth: Detecting prompt injection via semantic intent invariance.arXiv preprint arXiv:2508.20890, on PitharXiv
show 14 more
- [27] Shang Wang, Tianqing Zhu, Huajie Chen, Jiayang Li, Meng Yang, and Bo Liu. Hidden risks of Jev: An empirical study of security, privacy, and dual use.arXiv preprint arXiv:2610.04985, on PithPith's reviewarXiv
- [28] Tiantong Wu and Wei Yang Bryan Lim. Decision hijacking: Prompt injection attacks on Jev’s typed probabilistic decisions.arXiv preprint arXiv:2609.28613, on PithPith's reviewarXiv
- [29] Daniel Yang, Yao-Hung Hubert Tsai, and Makoto Yamada. On verbalized confidence scores for LLMs.arXiv preprint arXiv:2412.14737, on PithPith's reviewarXiv
- [32] uses a separate detector or a counterfactual reference; we read a cross-question invariant from the victim’s own answers. C Ways of counting hijacks All three readings of Section 3.2 use the same hijacks and 2000 resamples of articles.Pooled: each hijack is compared with every... on Pith
- [33] and of the fresh search run (searched text, fresh; Table 19), and four texts of the calibrated run (fresh; Table 21). F The choice of trace Table 16 gives the trace chosen on the development half of the fixed split and its AUROC on the test half. Each of the 200 redraws of Sec... on Pith
- [34] and Qwen3.8-Flash (served only through the API) are queried through the OpenAI-compatible endpoint of Alibaba Cloud Model Studio (Alibaba Cloud, 2026). The NLI encoder isdeberta-v3-large-zeroshot-v2.0(Laurer et al., 2023; He et al., 2023). Requests. gpt-6-luna receives the Cho... on Pith
- [2015] Yongxi Zhou, Junwei Yao, Yuanzhe Liu, Zihan Dong, Wenbo Ye, Jiaxi Wen, and Lai Yun Choi. Single canonical prompts underestimate LLM safety’s surface-form sensitivity.arXiv preprint arXiv:2608.02665, 2026a. Sci-FM Workshop at COLM 2026, non-archival. Yongxi Zhou, Wenbo Ye, Yuan... on PithPith's reviewarXiv
- [2017] doi: 10.1145/3128572.3140444. Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramèr. Membership inference attacks from first principles. InIEEE Symposium on Security and Privacy, on Pithdoi:10.1145/3128572.3140444
- [2020] doi: 10.18653/v1/2020.acl-main.503. Hankyeol Kim and Pilsung Kang. Same answer, different confidence: Protocol sensitivity in LLM confidence calibration.arXiv preprint arXiv:2605.27752, on PithPith's reviewarXiv
- [2022] Han Chen and Yingrui Li. Probability contracts: Accuracy, coherence, and decisions across LLM interfaces. arXiv preprint arXiv:2609.37470, on PithPith's reviewarXiv
- [2023] Seyedarmin Azizi, Erfan Baghaei Potraghloo, and Massoud Pedram. One word opens the gate: The option-channel attack on typed decision models as agent guardrails.arXiv preprint arXiv:2610.12292, on PithPith's reviewarXiv
- [2024] Qingcheng Zeng, Mingyu Jin, Qinkai Yu, Zhenting Wang, Wenyue Hua, et al. Uncertainty is fragile: Manipulating uncertainty in large language models.arXiv preprint arXiv:2407.11282, on PithPith's reviewarXiv
- [2025] doi: 10.18653/v1/2025.acl-long.890. Zheng Chen, Fei Yu, Haohao Huang, Yang Li, Anlong Chen, and Lei Chen. SecJev: Bringing security expertise to System One decision models.arXiv preprint arXiv:2610.03073, on PithPith's reviewarXiv
- [2026] Model catalog. Accessed 2026-10-12. Anastasios N. Angelopoulos and Stephen Bates. Conformal prediction: A gentle introduction.Foundations and Trends in Machine Learning, 16(4):494–591, 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