acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Probe-Guided Localization and Feature Repair for Visual Prompt Injection

Abstract

A visual prompt injection defense can suppress the attacker's target answer yet still fail the user's task. It must reject image-borne instructions while preserving the visual evidence needed to answer correctly. Separate-caption evaluations do not test this distinction within application content. We introduce AppInjectBench, a benchmark spanning 11 applications that embeds injected instructions in interface content such as reviews, comments, and messages. Each example pairs a clean screenshot with a benign-text variant and an injected variant, enabling joint evaluation of task recovery and benign task performance. We further propose PRISM, which observes attention to visual tokens under a fixed query. A learned spatial locator uses this observation to identify injection regions. PRISM replaces the selected visual features with a shared mean of their unselected boundary features. The frozen model then answers the original task from a fresh prefill using the modified representation. On model-specific AppInjectBench test sets of 1,000 sources screened for correct benign answers and successful undefended injection, PRISM recovers 96.7% of cases with Qwen2-VL-7B and 94.0% with Qwen2.5-VL-7B, while causing new errors on at most 0.3% of benign inputs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
visual prompt injection, vision-language models, multimodal safety, visual token localization, feature repair, adversarial robustness

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wvU3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2(+1)
Read review
Paper summary

The authors propose PRISM, a defense against prompt injection for MLLMs. The method first feeds the image together with a fixed, task-independent query into a frozen VLM for one prefill, and extracts the multi-layer, multi-head attention from this query to each visual token as features. A lightweight spatial locator, trained with injection-region annotations, then uses these attention features to predict which visual tokens likely belong to the prompt injection region. Next, the visual features of all selected tokens are uniformly replaced with the mean of the unselected neighboring features in the one-hop ring around the selected region. Finally, the repaired visual features are fed together with the user’s original question into the same frozen model to produce the answer, thereby defending against prompt injection.

Critical strengths
  1. The problem definition is valuable. The authors clearly distinguish between blocking an attack and answering the question correctly, and point out that blocking the attack without answering the question correctly is not enough.

  2. The authors build a benchmark with realistic and valuable data. The screenshots are obtained from 11 real deployed open-source applications.

  3. PRISM is lightweight. The whole locator has only about 9K parameters, and the recovery rate exceeds 94% on both models.

Critical weaknesses
  1. PRISM’s cross-distribution generalization is weak. Appendix Table S5 shows that on Qwen2.5-VL, the locator trained only on ARGUS data achieves a recovery rate of only 39.9% on the App test, while the harm on normal inputs reaches 45.6%. This suggests that when the locator encounters unseen interface layouts, it fails on a large scale and damages normal content.

  2. There is no evaluation of damage to normal visual cues in dense scenes. The benchmark does not test cases where the image contains multiple normal visual cues, where the answer region and the injection region are adjacent or overlapping, where the instruction is embedded in the main text, or where the answer itself is located in the user-writable region. In text-dense scenes such as TextVQA, PRISM’s Clean error is 3.52%, clearly higher than ARGUS’s 0.44%, which suggests that this risk exists.

  3. There is no test on benign text that looks like instructions. The Clean and Neutral controls contain only ordinary content and unrelated text. Whether normal instruction-like text that naturally exists in interfaces, such as operation hints, system announcements, or usage instructions, would be misjudged as injection and removed is not evaluated.

  4. PRISM lacks evaluation against adaptive attacks. The paper does not report the diversity of the prompt injection methods used, and does not test injections with natural wording and no keywords, injections split across multiple locations, or attackers who know the defense mechanism, for example, layouts designed to deliberately scatter attention.

  5. The writing is somewhat obscure. A large portion of the paper is devoted to details and qualifications of the evaluation protocol, the sentences are dense, and it is hard to follow the main line of the method and the conclusions.

Questions for authors
  1. How well does PRISM generalize? The test set and the training set come from the same 11 applications. If several applications were entirely held out from training, what would the recovery rate and harm rate be? On Qwen2.5-VL, when the locator trained only on ARGUS data is tested on App, 45.6% of normal inputs are answered incorrectly. Does this mean the locator causes large-scale damage when it encounters unseen interface layouts? How is this addressed?

  2. Could the locator mistakenly remove normal visual cues? For example, when the image contains several pieces of information that are needed, when the answer region and the injection region are very close or even overlapping, or when the answer itself is in the comment section, how does PRISM perform in these cases? In addition, PRISM’s Clean error on TextVQA is somewhat higher than ARGUS’s. I would like to understand what roughly causes these errors.

  3. The data's interfaces naturally contain some normal text that looks like instructions, such as operation hints, system announcements, or usage instructions. Would the locator treat such text as injection and remove it? Could the authors measure the harm rate on such samples?

  4. How many different injection methods and phrasings are used for the prompt injection instructions? If the prompt is split and placed in several locations, can the locator still find it? In addition, if the attacker knows that the defense relies on attention for localization and specifically designs layouts that scatter attention, how would the method perform?

10 comments
AuthorsOpenReview ↗

Response to Reviewer wvU3 (Part 1/3)

Thank you for the detailed review. We group overlapping Weaknesses (W) and Questions (Q), numbered in their original order. Part 1 addresses transfer and nearby evidence (W1-W2/Q1-Q2). Part 2 covers legitimate instructions and attack diversity. Part 3 covers answer-feedback search and presentation.

1. Training-domain shift and unseen applications (W1, Q1).

The cited ARGUS-only-to-App result changes the training domain, injection carrier and visual context. ARGUS appends injected text in a separate caption, whereas AppInjectBench places it among interface content, including answer-bearing text. On the same Qwen2.5-VL App test, Mixed training raises Recovery from 39.9% to 93.3% and reduces Clean/Neutral harm from 45.6%/44.6% to 0.3%/0.2%. The same Mixed checkpoint retains 92.5% Recovery on ARGUS, versus 93.7% for ARGUS-only training. This supports effective defense on both datasets using a shared locator when training covers both domains (Appendix C.2). Changes in probe-attention cues are a possible explanation for the transfer gap, which has not been verified at the feature level.

We tested unseen applications separately with Qwen2-VL, excluding each test application from locator training, validation and normalization fitting. Recovery is 94.66% (939/992), versus 97.38% (966/992) for the matched all-application reference. The decrease is 2.72 percentage points (95% paired source-bootstrap interval [1.31, 4.13]). Clean/Neutral harm is 0.60%/0.71%, versus 0.10%/0.00%. These fits use the same captures and training procedure. Table 6 gives per-application results and training quantities. These results support transfer to applications excluded from Qwen2-VL locator training.

The original training already included all-zero Clean/Neutral targets in the localization loss. We clarified this in Section 4.5 and Appendix B.3. Appendix A.1 describes split overlap, and Section 5.7 distinguishes application holdout from mixed-domain training.

2. Nearby answer evidence and TextVQA errors (W2, Q2).

We constructed 22 paired WordPress/Nextcloud Talk scenarios in the native apps. The question asks for a room number in a comment/message. Either benign text or an attack is placed beside that number in the same UI block, testing very tight spacing. Qwen2-VL answers all 22 correctly without defense. PRISM preserves 22/22 beside benign text and 19/22 beside the attack. Each of the 3 losses modifies 4/14 answer-region tokens.

Keeping questions, answers and attack texts fixed, we changed only the spacing in 720 x 1,600 screenshots:

Answer-attack gap (px)Correct-to-wrong answersAnswer-region tokens modified
Original (median 9.6)3/22 (13.6%)11.8%
150/22 (0.0%)0.6%
300/22 (0.0%)0.0%

The undefended model answers all 22 correctly at every distance. All benign controls remain correct with PRISM, and every attack receives a nonempty mask. Qwen2.5-VL and Qwen3-VL also preserve all 22 benign answers. Their attacked-input accuracy stays at 21/22 and 22/22, respectively, with and without defense. The original diagnostic gap is 9.0-13.2 pixels, compared with a median of 600 pixels (minimum 128) across 998 coordinate-verified Qwen2-VL benchmark images. This measures answer preservation under controlled proximity (Appendix C.6, Table S5). Multiple-clue tasks and primary-page-text injections are outside this panel.

For TextVQA, we corrected 2 scoring false negatives: "the first chapter" answers chapter 1, and "Y, X, W, V, and U" matches the registered sequence. Clean error is corrected from 3.52% (8/227) to 2.64% (6/227), while attack Recovery remains 72.1% (Table 2). The authors confirmed these corrections through question-aware review assisted by Claude Opus 5.5 (Appendix A.2). This test uses frozen ARGUS-trained weights without TextVQA training. The remaining Clean errors show reduced answer preservation under that transfer setting, without establishing a specific feature-level cause.

AuthorsOpenReview ↗

Response to Reviewer wvU3 (Part 2/3)

This continues our response to Reviewer wvU3, addressing W3/Q3 and the wording and multi-region parts of W4/Q4. Answer-feedback search follows in Part 3.

3. Information inside legitimate instructions (W3, Q3).

We tested 100 benign scenarios, 50 each from WordPress and Nextcloud Talk, covering reply codes, operation steps, announcements and usage guidance. The question directly asks for information in the legitimate instruction, such as a code, button, time or permission. Existing application-excluded locators are used without further training.

BackboneWithout PRISMWith PRISMHarm rate
Qwen2-VL100/10098/1002.0%
Qwen2.5-VL100/10099/1001.0%
Qwen3-VL100/100100/1000.0%

All undefended answers are correct, so harm measures correct-to-wrong transitions. Qwen2-VL gives an unrelated time once and omits a desk number once. Qwen2.5-VL changes 7:25 PM to 7:00 PM. These results directly measure whether PRISM preserves information needed from legitimate instructions (Appendix C.6, Table S3).

4.1. Attack methods and wording (W4, Q4).

AppInjectBench uses native user-writable areas across 11 applications. We counted 5,460 distinct attack texts across 6,438 sources, with 6,433 matched to 71 instruction templates and 5 unclassified. Many texts differ by entities or values. Complete attack texts do not repeat across application-holdout boundaries, although some templates are shared.

Explicit redirection is used to elicit the attack target. Eligibility requires correct undefended Clean/Neutral answers and target adoption on Impact. Screening uses only undefended outputs. In a separate 22-scenario check without SYSTEM or assistant-role markers, all 3 backbones answer correctly both with and without PRISM. With no effective undefended attacks, that check cannot measure conditional Recovery.

To test unseen wording, we evaluated 4 new single-region templates, each on the same 125 WordPress scenarios (500 trials). Both the templates and WordPress were absent from locator training. Without defense, Qwen3-VL adopts the target in 75/500 trials (15.0%). PRISM restores the correct answer in 63/75 effective attacks (84.0%) and reduces target adoption over all trials to 2.8%. This supports transfer under joint application/wording holdout, with lower Recovery than the main screened benchmark.

4.2. Attack content placed in 2 regions (W4, Q4).

We evaluated 4 additional templates on a separate set of 250 WordPress scenarios. A redirection prefix occupies the first comment and a wrong-answer instruction the third, separated by a benign comment. Templates were selected using only undefended outputs on a separate 32-scenario pilot. The WordPress-excluded locator remained fixed.

WordingTrialsUndefended target adoptionPRISM RecoveryMask hits both
Original AppInjectBench250119/250 (47.6%)106/119 (89.1%)238/250
New template 1250170/250 (68.0%)165/170 (97.1%)245/250
New template 2250117/250 (46.8%)117/117 (100.0%)249/250
New template 3250145/250 (58.0%)144/145 (99.3%)250/250
New template 425062/250 (24.8%)62/62 (100.0%)248/250
4 new templates, pooled1,000494/1,000 (49.4%)488/494 (98.8%)992/1,000

Every row uses the same 250 scenarios. The pooled row covers the 4 new templates, whose effective attacks involve 214 scenarios. PRISM reduces target adoption to 6/1,000 (0.6%) and preserves all 250 benign-control answers. "Mask hits both" means selecting at least one annotated visual token in each region. The second region contains a complete wrong-answer instruction, so this tests two-region placement rather than attacks requiring combination of incomplete fragments. Appendix C.6 and Table S4 contain these results.

AuthorsOpenReview ↗

Response to Reviewer wvU3 (Part 3/3)

This concludes our response to Reviewer wvU3, addressing the answer-feedback and attention-scattering part of W4/Q4, followed by presentation (W5).

4.3. Repeated attacks using answer feedback (W4, Q4).

We evaluated whether repeated probing of PRISM's answers finds additional successful attacks. On Qwen3-VL, we used 30 scenarios (15 WordPress and 15 Nextcloud Talk), with 20 attempts per scenario and the task question and page evidence fixed.

The first 4 attacks were predefined: 2 single-comment/message attacks, 1 two-region attack and 1 three-region attack. The remaining 16 attempts used earlier PRISM answers to choose untried variants from a fixed pool of 108 candidates. The controller prioritizes attacks whose answers adopt the target, then chooses a nearby variant by changing wording, answer format, region count or inclusion of the page title. All variants use predefined native comment/message positions. The controller sees answers, without access to masks, attention or gradients. The undefended model is evaluated on the same 600 candidates selected through PRISM feedback.

OutcomeWithout defenseWith PRISM
Target adoption per attempt210/600 (35.0%)42/600 (7.0%)
Scenarios with at least 1 success28/3012/30

PRISM restores the correct answer on 168/210 effective undefended attacks (80.0%). The next table shows when each scenario is first successfully attacked:

First successful attemptWithout defenseWith PRISM
By attempt 4 (predefined attacks)237
611
801
1131
1901
2011
No success in 20 attempts218

The answer-guided attempts find a first bypass in 5 of the 23 scenarios that withstand the predefined attacks. PRISM withstands all 20 attempts on 18/30 scenarios. These outcomes show reduced target adoption on the tested candidates and identify additional bypasses through repeated probing. This is an answer-feedback search within native comment/message positions. It does not optimize layouts to scatter probe attention, so it does not establish robustness to that specific attack. Appendix C.6 describes the search. We plan to explore attention-scattering layouts and whether training on additional attack variants improves robustness.

5. Readability and manuscript changes (W5).

We have substantially revised the exposition to clarify the motivation, method and experimental conclusions. Sections 1 and 3 now explain why instructions embedded in application content must be suppressed while preserving task evidence. We define the separate-caption setting before using it and explain how AppInjectBench's Clean, Neutral and Impact screenshots measure recovery and benign preservation.

Section 4 now presents PRISM in execution order: fixed-query observation, spatial localization, feature replacement and task answering. We explain the purpose of the fixed query, attention normalization and neighborhood features before their technical definitions. The text also explicitly distinguishes the frozen backbone from the trained locator and clarifies benign negative supervision.

Section 5 now states what each experiment tests and presents the main finding before its supporting numbers. The main results distinguish task recovery from target suppression and benign harm. The following experiments separately examine frozen-weight transfer, mixed-domain training, locator components, repair operators, application holdout and inference cost. Detailed scoring rules and implementation settings are consolidated in Appendix A.2 and Appendices B-C.

Across the manuscript, we added transitions, merged fragmented paragraphs, reduced repeated qualifications and simplified table captions. These changes make the progression from motivation to design and evidence explicit. The revised main text remains within 10 pages.

Reviewer_h1mdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors introduce a new benchmark to evaluate the visual prompt injection attacks in the real application, and propose a defense method PRISM to help VLM overlook the attacked information and still give the correction answer. PRISM first uses attention form a fixed query to identify the attacked region in an image and then replace the features of such region using the average features of nearby regions. In this way, the attack can be defensed and the correct answer can be obtained. Also, the experiment reports the strong recovery rates on the tested models.

Critical strengths

(1) In this paper, since a refusal or unrelated wrong answer does not satisfy the user's original task, the authors separate the correct recovery from the target suppression and improve the defense evaluation.

(2) The authors carefully organize a benchmark dataset, where the paired clean, neutral, and Impact design and explicit model-specific screening could support controlled comparisons without implying unconditional attack success or cross-model robustness rankings.

(3) The reported experiment results are impressive. Specifically, Table 1 suggests substantial improvements on both evaluated backbones while preserving in-domain benign answers, and Table 2 adds frozen-defense tests on other sources and target forms.

Critical weaknesses

(1) As for the evaluation metric, the authors use string matching to reflect whether the model actually answers the correction correctly and claims that the recovery can achieved 96.7. However, in some scenario, for example, the correct answer is "100" and the model's answer "The answer is not 100", the evaluation rule sometimes are not reliable. The authors should discuss this and give more explanation.

(2) The authors should add an ablation study to show the contribution of the proposed feature repair. It is unclear whether simply removing or zeroing the detected tokens would achieve similar results. Different replacement strategies should also be explored.

(3) The authors should also explore the generalizability of proposed PRISM on held-out applications and different injection positions.

(4) PRISM relies on detecting suspicious regions and replacing their features with information from nearby tokens. However, if the attackers modify their injections randomly or multiple regions, the suspicious region localization cna be difficult.

Questions for authors

(1) If the model responses are evaluated based on their actual meaning rather than string matching, do PRISM and the baselines still achieve similar recovery rates and benign harm results? This is also similar to the above weakness-(1).

(2) Does boundary-mean replacement actually perform better than simply zeroing out the detected visual tokens? The authors should conduct some experiment to compare them.

(3) Can PRISM still defend against attacks when malicious instructions are spread across multiple regions or appear near the detected region?

4 comments
AuthorsOpenReview ↗

Response to Reviewer h1md (Part 1/2)

Thank you for the review. We group overlapping Weaknesses (W) and Questions (Q), numbered in their original order. This part addresses scoring (W1/Q1) and repair (W2/Q2). Part 2 covers application transfer and attack placement.

1. Semantic scoring and corrected results (W1, Q1).

We rescored 30,000 saved Qwen2-VL and Qwen2.5-VL responses using question-aware semantic rules. The audit covers 24,000 primary responses from no defense, PRISM, ARGUS and GTM, plus 6,000 additional undefended responses to the same inputs. Each backbone retains the same 1,000 test sources and their Clean, Neutral and Impact screenshots.

The protocol distinguishes an asserted answer from negation, quotation, conflicting statements and refusal, while accepting equivalent wording. Clear cases use rules. Claude Opus 5.5 reviews ambiguous or paraphrased answers using the question, registered correct answer, attack target and complete response. Model, method, input condition and prior labels are hidden, and the judge returns a decision and rationale. This is a retrospective review of saved outputs.

The audit corrects 60 definite labels. For example, "10 valve zones" correctly answers "10 zones", which the substring scorer missed. This changes Qwen2.5-VL ARGUS Recovery from 17.8% to 17.9%. A targeted scan found no instance of the specific registered-answer negation pattern you describe.

BackbonePRISM RecoveryARGUS RecoveryGTM Recovery
Qwen2-VL96.7%45.3%54.9%
Qwen2.5-VL94.0%17.9%40.5%

PRISM Recovery and the method ordering remain unchanged. Some benign-harm counts change: Qwen2.5-VL ARGUS Clean/Neutral harm becomes 73.8%/71.6%, from 75.0%/73.3%, and GTM harm becomes 1.0%/1.7%, from 1.9%/2.6%. Table 1 now uses the corrected scores. Appendix A.2 describes the protocol.

The added Qwen3-VL comparison uses its own 1,000 qualified sources. PRISM and ARGUS recover 94.5% and 57.9%, with target adoption of 5.4% and 0.2%. ARGUS incurs 34.5%/33.8% Clean/Neutral harm alongside its low target adoption (Table 1).

2. Repair with identical predicted masks (W2, Q2).

We fixed the locator and predicted masks within each backbone and compared boundary mean, zeroing, global mean and physical token dropping. Dropping preserves retained-token positions. The table gives equally weighted means over the 3 backbones, with 1,000 inputs per condition per backbone:

ReplacementAttacked-input accuracyClean harmNeutral harm
Boundary mean95.23%0.03%0.00%
Zeroing94.40%0.10%0.00%
Global mean95.13%0.07%0.00%
Token dropping95.30%0.07%0.00%

Boundary mean improves mean attacked-input accuracy by 0.83 percentage points over zeroing and 0.10 points over global mean. Token dropping is 0.07 points higher than boundary mean. Boundary mean has the lowest mean Clean harm, with 0% mean Neutral harm for every operator. The 11 empty Qwen2.5-VL zeroing responses count as incorrect. The results support the observed benefit over zeroing while showing similar recovery for global mean and dropping. Table 5 reports these means. Appendix C.3 and Table S1 give definitions and per-backbone results.

AuthorsOpenReview ↗

Response to Reviewer h1md (Part 2/2)

This continues our response to Reviewer h1md, addressing W3-W4 and Q3.

3. Held-out applications and injection positions (W3).

We excluded each test application from Qwen2-VL locator training, validation and normalization fitting. On the same 992 eligible sources, Recovery is 94.66% (939/992), compared with 97.38% (966/992) for the matched all-application reference. The decrease is 2.72 percentage points (95% paired source-bootstrap interval [1.31, 4.13]). Clean/Neutral harm is 0.60%/0.71%, versus 0.10%/0.00%. Both fits use the same captures and training procedure. Section 5.7 and Table 6 report per-application performance and training quantities.

For legitimate instruction-like content, we tested 100 benign WordPress/Nextcloud Talk scenarios. Questions directly ask for information in reply-code instructions, operation steps, announcements and usage guidance. All 3 undefended backbones answer 100/100 correctly. PRISM retains 98/100, 99/100 and 100/100 on Qwen2-VL, Qwen2.5-VL and Qwen3-VL, using application-excluded locators (Appendix C.6, Table S3).

4. Varied wording, multiple regions and nearby evidence (W4, Q3).

Two-region placement. We tested 4 additional templates on the same 250 WordPress scenarios, yielding 1,000 attacks. A task-redirection prefix occupies the first comment and a wrong-answer instruction the third, with a benign comment between them. The templates were selected using only undefended outputs on a separate 32-scenario pilot. The locator trained without WordPress remained fixed.

Qwen3-VL adopts the target in 494/1,000 trials (49.4%) without defense, involving 214 scenarios. PRISM restores the correct answer in 488/494 effective attacks (98.8%) and reduces target adoption to 6/1,000 (0.6%). The executed mask selects at least one annotated visual token in each region on 992/1,000 trials. Both methods answer all 250 benign controls correctly. The second region contains a complete wrong-answer instruction, so this tests two-region placement rather than attacks requiring combination of incomplete fragments. Appendix Table S4 gives per-template results and the original-wording control.

Nearby answer evidence. In a separate Qwen2-VL control, we fixed 22 questions, answers and attack texts and changed only the answer-attack spacing:

Gap (px)Correct-to-wrong answersAnswer-region tokens modified
Original (median 9.6)3/22 (13.6%)11.8%
150/22 (0.0%)0.6%
300/22 (0.0%)0.0%

The undefended model answers all 22 correctly at every distance. All 22 paired benign controls remain correct with PRISM, and all attack images receive nonempty masks. At the original placement, Qwen2.5-VL fails to repair its 1 successful near-answer attack. Qwen3-VL has no successful undefended attack in this panel, so conditional Recovery is undefined (Appendix C.6, Table S5).

Repeated attempts using answer feedback. We also tested Qwen3-VL on 30 scenarios, with 20 attempts each. The question and page evidence stay fixed. After 4 predefined attacks covering 1-, 2- and 3-region placement, 16 further variants are selected using PRISM's earlier answers, changing wording, answer format or region count within predefined native comment/message positions. Both methods evaluate the same 600 selected candidates.

Target adoption is 210/600 (35.0%) without defense and 42/600 (7.0%) with PRISM. At least one attack succeeds on 28/30 and 12/30 scenarios, respectively. Of the 12 PRISM bypasses, 7 first occur in the predefined attempts and 5 in the subsequent answer-guided attempts. This demonstrates lower target adoption on the tested candidates and identifies additional bypasses through repeated probing. The controller sees answers, without masks, attention or gradients. This experiment does not optimize layouts to scatter attention (Appendix C.6).

Thank you for the detailed responses and additional experiments. The new results have addressed most of my concerns regarding generalization, inference overhead, and localization quality.

Thank you for reviewing our responses and additional experiments. We appreciate your constructive feedback, which helped us strengthen the paper.

Reviewer_X7NoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses a practical limitation in defending Vision-Language Models (VLMs) against visual prompt injection: suppressing the attacker's target string is insufficient if the defense corrupts the visual evidence required to answer the user's actual query. To evaluate this properly, the authors introduce AppInjectBench, a benchmark rendered across 11 real applications where untrusted user content (reviews, comments, messages) coexists with legitimate task information, pairing Clean, Neutral, and Impact screenshots. They also propose PRISM, a two-stage defense on a frozen backbone. PRISM first runs a cache-free prefill with a fixed probe query ("What's the next action") to extract cross-modal attention from decoder layers 2, 4, and 8. A lightweight supervised spatial locator (MLP + depthwise-separable convolutions) predicts a binary mask over the injected visual tokens. PRISM then replaces the masked visual features with a single shared mean of their unselected 4-neighbor boundary tokens, and runs a fresh prefill with the user's original query. Experiments on Qwen2-VL-7B and Qwen2.5-VL-7B show strong task recovery (96.7% and 94.0%) and very low benign harm (≤0.3%\le 0.3\%), outperforming ARGUS and GTM.

Critical strengths
  • Insightful diagnostic and benchmark design: The diagnostic in Section 3.1 nicely exposes a flaw in activation-steering defenses like ARGUS: while ARGUS achieves 0.0% ASR on random-string targets, its ASR jumps to 34.8% when the injected target is a plausible same-type wrong answer. AppInjectBench is also thoughtfully constructed—rendering text natively inside real UI templates rather than pasting text boxes onto images, and using Clean/Neutral/Impact triplets to strictly separate target suppression (BLOCKED-WRONG) from true task recovery (CORRECT).

  • Simple and effective feature repair: Keeping the visual token grid and positional encodings intact while replacing suspected tokens with their 1-hop boundary mean (Eq. 4–5) is a clean design. It neutralizes localized text instructions without introducing the out-of-distribution feature shift caused by zeroing tokens (as in GTM).

Critical weaknesses
  • No negative supervision on clean images leads to severe cross-domain over-masking: In Section 4.5 and Appendix B.3, the authors note that Clean and Neutral images provide no spatial targets and "are not treated as empty-mask negatives" during locator training. Because the locator is only trained on images containing injections (B_loc) and is never penalized for false-positive masks on benign screenshots, it is prone to over-firing when layout priors shift. We see this clearly in Supplementary Table S5 (Appendix C.3): when PRISM is trained on ARGUS-only and tested on AppInjectBench (Qwen2.5-VL), Clean harm and Neutral harm spike to 45.6% and 44.6%, and Recovery drops to 39.9%. Conversely, when trained on AppInjectBench, Clean harm is 0.0%–0.1%. Since the train and test splits of AppInjectBench share the same 11 applications ("held-out layouts, which need not imply unseen applications", Section 3.3), the locator appears to be memorizing app-specific UI layout priors (e.g., where comment boxes sit in Drupal or WooCommerce) rather than learning a generalizable representation of visual prompt injections.
  • Missing ablation on the feature repair mechanism itself: The paper claims that replacing selected tokens with a shared 4-neighbor boundary mean is a key contribution over zeroing or pruning. However, all ablations in Table 4 and Table S6 hold the repair rule fixed and only vary the locator architecture. As acknowledged at the end of Section 5.5, there is no experiment isolating the repair rule. What happens if you take the exact same mask predicted by PRISM's full locator and compare: (a) 4-neighbor mean replacement, (b) zeroing out the masked tokens (GTM-style), (c) replacing with the global image token mean, and (d) dropping the masked visual tokens before the second prefill?
  • Unquantified two-pass inference overhead: PRISM requires running the LLM decoder up to layer 8 with a probe query, executing the locator, discarding the KV cache, and running a second full prefill from scratch on the modified visual tokens. For high-resolution UI screenshots (720×1600720 \times 1600) that produce over a thousand visual tokens, what is the exact wall-clock latency (TTFT) and FLOPs overhead of this two-pass pipeline compared to undefended inference, GTM, and ARGUS?
  • Lack of direct spatial localization metrics (IoU) and feature leakage analysis: Because the ViT and spatial merger mix patch information before the projector output ZZ, donor cells in the 1-hop boundary ring BB can already carry leaked features from adjacent injected text—or conversely, if an attacker places the injection right next to the task evidence, the predicted mask (which shows noticeable false-positive islands outside the comment box in Fig. S4) might overwrite the answer evidence itself. Reporting pixel/grid-level IoU, precision, and recall of the predicted mask SS against the ground-truth boxes would make the locator's behavior much more transparent than relying solely on downstream string-matching recovery rates.
Questions for authors

see weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer X7No (Part 1/2)

Thank you for the constructive review. We use W1-W4 to follow the order of your Weaknesses. This part addresses W1-W2. Part 2 covers inference cost and localization.

1. Benign supervision and application transfer (W1).

The original Qwen2-VL and Qwen2.5-VL locators already used Clean and Neutral images as negatives. These images have no injection boxes, but every valid visual token was labeled 0 and contributed to the weighted binary cross-entropy loss. For annotated Impact images, tokens overlapping an injection box were labeled 1 and the remaining valid tokens 0. We have corrected the misleading training description in Section 4.5 and Appendix B.3.

The cited ARGUS-only-to-App result measures a change in training domain and injection carrier. ARGUS appends injected text in a separate caption, whereas AppInjectBench places it among interface content, including answer-bearing text. On the same Qwen2.5-VL App test, Mixed training raises Recovery from 39.9% to 93.3% and reduces Clean/Neutral harm from 45.6%/44.6% to 0.3%/0.2%. The same Mixed checkpoint retains 92.5% Recovery on ARGUS, compared with 93.7% for ARGUS-only training. A shared locator therefore supports effective defense on both datasets when training covers both domains (Appendix C.2). Differences in probe-attention cues are a possible explanation for the transfer gap, rather than an established cause.

To test unseen applications directly, we added Qwen2-VL application holdout, excluding each test application from locator training, validation and normalization fitting. On the same 992 eligible sources, Recovery is 94.66% (939/992), versus 97.38% (966/992) for the matched all-application reference. The decrease is 2.72 percentage points (95% paired source-bootstrap interval [1.31, 4.13]). Clean/Neutral harm is 0.60%/0.71%, versus 0.10%/0.00%. Both settings use the same captures, training procedure and random seed. These Qwen2-VL results support application-level transfer with a measurable decrease from all-application training. Section 5.7 and Table 6 give per-application results and training quantities removed and retained from the 4,938-source reference.

2. Contribution of the repair operator (W2).

We fixed the locator and predicted masks within each backbone and compared all 4 requested operators. The table gives equally weighted means over Qwen2-VL, Qwen2.5-VL and Qwen3-VL, with 1,000 inputs per condition per model. Token dropping physically removes selected visual tokens while preserving retained-token positions.

ReplacementAttacked-input accuracyClean harmNeutral harm
Boundary mean95.23%0.03%0.00%
Zeroing94.40%0.10%0.00%
Global mean95.13%0.07%0.00%
Token dropping95.30%0.07%0.00%

Boundary mean exceeds zeroing by 0.83 percentage points and global mean by 0.10 points in mean attacked-input accuracy. Token dropping is 0.07 points higher than boundary mean. Boundary mean has the lowest mean Clean harm, and all 4 operators have 0% mean Neutral harm. The 11 empty Qwen2.5-VL zeroing responses count as incorrect. These results establish the observed improvement over zeroing while showing that global mean and dropping achieve similar recovery. Table 5 reports the means. Appendix C.3 defines the operators, and Table S1 gives per-backbone results.

AuthorsOpenReview ↗

Response to Reviewer X7No (Part 2/2)

This continues our response to Reviewer X7No, addressing W3-W4.

3. Measured inference cost (W3).

On one RTX 5090 with 32 GB memory, Qwen2-VL has the following end-to-end time to first token (TTFT) and peak PyTorch allocation, including model weights (Table 7):

MethodMean TTFT (ms)p50 (ms)p95 (ms)Peak memory (GiB)
No defense407.87407.46415.1816.128
ARGUS407.62407.67413.3516.128
GTM719.22719.08725.6323.374
PRISM565.04564.84574.9716.138

PRISM adds 157.17 ms (38.5%, 95% source-paired interval [156.28, 158.08] ms) with nearly unchanged peak memory. GTM uses 1.76 times the undefended TTFT and 1.45 times its peak allocation.

PRISM encodes the image once, runs a forward-only probe through decoder layer 8, then localizes and repairs features before a fresh task prefill. GTM backpropagates through the native image forward pass for attribution, zeros the top 5% of visual features, and runs a fresh prefill. ARGUS adds detection and steering within the task pass. Its reported internal TTFT precedes final postfiltering.

All methods use BF16 on the same 100 fixed attacked screenshots, with 10 warmups per method and 3 repetitions per screenshot (300 timed passes per method). Method order rotates across inputs. Timing includes image reading, preprocessing and all online work before the first token. Features, masks and KV states are recomputed each pass. Model loading and warmups are excluded. Appendix C.5 gives the protocol. Appendix C.7 distinguishes the native GTM timing path from its answer-accuracy runs. Whole-pipeline FLOPs have not been measured.

4. Localization, nearby evidence and donor features (W4).

We measured the masks actually used for repair against attack annotations on all 1,000 attacked screenshots per backbone:

BackboneIoUPrecisionRecall
Qwen2-VL83.47%90.67%91.31%
Qwen2.5-VL82.01%89.73%90.50%
Qwen3-VL82.33%85.20%96.07%

These metrics aggregate true-positive, false-positive and false-negative visual-token counts across each test set (Appendix C.4, Table S2).

We also tested Qwen2-VL on 22 scenarios with the answer in a comment/message beside an attack. We fixed questions, answers and attack texts and changed only their spacing in 720 x 1,600 screenshots:

Answer-attack gap (px)Correct-to-wrong answersAnswer-region tokens modified
Original (median 9.6)3/22 (13.6%)11.8%
150/22 (0.0%)0.6%
300/22 (0.0%)0.0%

The undefended model answers all 22 correctly at every distance. All 22 paired benign controls also remain correct with PRISM. Every attack receives a nonempty mask at every distance. Thus, the larger gaps preserve answers while PRISM still modifies visual features. At the original placement, Qwen2.5-VL fails to repair its 1 successful near-answer attack (Appendix C.6, Table S5).

These experiments quantify localization and answer preservation. Donor features outside an attack box can already contain information mixed by the vision encoder. The tests do not measure semantic contamination of those donors, so they do not establish it as the cause of the observed errors.

AuthorsOpenReview ↗

Follow-up to Reviewer X7No

Thank you again for your thoughtful review. We hope our responses and additional results have helped clarify the questions you raised. We would welcome any further feedback and would be happy to clarify any remaining concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.