Escape Meridian

Xing, Jiawei, Wang, Li, Meng, Xiangtao, Zang, Chuanchao, Wen, Rui, Li, Zheng · Nov 19, 2026

arXiv cs.AI: ChronoPatch Black-Box Attacks on Vision-Language Models

Markdown · JSON · Original source

The authors of ChronoPatch describe a black-box attack on vision-language models (VLMs) that uses only the final response to prolong generation. They argue this matters because response length can determine and computational cost in generative inference services, creating an availability risk even when an attacker cannot see gradients, logits, or token-level feedback. Their method finds visual delay anchors through response-length-guided evolutionary search, then reduces perturbation magnitude through subspace-calibrated decision-based optimization. Across four open-source VLMs, the authors report increased inference latency and energy consumption under a 5% spatial budget, alongside more repetitive outputs. On three commercial VLM , they report longer responses and higher client-observed latency with query budgets of 500 per API. These are reported experimental results, not evidence here of effects across all deployed systems. The authors say the findings motivate generation-aware safeguards; the abstract does not specify which safeguards would work.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

arXiv cs.AIArticle · arxiv.orgOpen original source

Original source

cs.AI · Article

Source details

Original source date: Nov 19, 2026 (2026-11-19).

Date basis: Original submission date in the publisher’s citation record.

Added to Escape Meridian: 2026-11-20T09:01:36.993Z. First observed: 2026-11-20T05:07:49.630Z. Retrieved: 2026-11-20T09:00:52.223Z.

Referenced by

Recorded citations in public readings. A reference alone does not establish agreement or influence.

No public references recorded here yet.

Reference data (JSON)

Limits

The supplied evidence is the abstract and page text, not the paper's detailed methods or experimental analysis; it does not identify the four open-source models or three commercial . The reported results concern the tested models and API query budgets; the supplied text does not establish how broadly they generalize or which generation-aware safeguards would work.

Open question

What generation-aware safeguards, if any, would address the final-response-only availability risk described by the authors?

Source-stated

The authors present ChronoPatch as a black-box attack that prolongs VLM generation using only the final response.

We present ChronoPatch, a black-box attack that induces prolonged VLM generation requiring only the final response.

Source-stated

The authors frame prolonged generation as an availability risk because response length affects latency and computational cost.

autoregressive decoding makes response length a key determinant of latency and computational cost.

Source-stated

The method searches for visual delay anchors and then minimizes perturbation magnitude.

It discovers visual delay anchors through response-length-guided evolutionary search, then minimizes perturbation magnitude through subspace-calibrated decision-based optimization.

Source-stated

Across four open-source VLMs, the authors report increased latency and GPU energy consumption under a 5% spatial budget.

Across four open-source VLMs, ChronoPatch increases inference latency by 47.86%--293.34% and GPU energy consumption by 41.44%--249.70% under a 5% spatial budget.

Source-stated

On three commercial VLM APIs, the authors report increased response length and client-observed latency with query budgets of 500 per API.

On three commercial VLM APIs, ChronoPatch increases response length by 37.6%--81.5% and client-observed latency by 1.82x--2.29x under query budgets of 500 per API.

Source-stated

The authors say their findings motivate generation-aware safeguards for deployed generative AI services.

These results reveal a generation-based availability risk that persists under final-response-only access and motivate generation-aware safeguards for deployed generative AI services.

Source and contributor context

Xing, Jiawei, Wang, Li, Meng, Xiangtao, Zang, Chuanchao, Wen, Rui, Li, Zheng · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Latency

The delay between sending a request and receiving the result.

Graphics processing unit

A chip that performs many calculations at once; the main hardware used to train and run AI models.

API

An application programming interface: a defined way for software to request data or actions from another system.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Editorial provenance

Version 1. Publication ID: 06e4f1cd-39dc-44e5-8357-511ef00cbb7b.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-20T09:01:36.993Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 6183064436e0911856ae0068b3865bd4909e2bb7a0ff1be583edbf7e32afca96. Source text hash: 30141bedd111b284396a118959848e5f44fc00b7b1308d52f5475cb8045198cb.

Revision and source-check history

Version 1 · 2026-11-20T09:01:36.993Z

arXiv cs.AI: ChronoPatch Black-Box Attacks on Vision-Language Models

The authors of ChronoPatch describe a black-box attack on vision-language models (VLMs) that uses only the final response to prolong generation. They argue this matters because response length can determine and computational cost in generative inference services, creating an availability risk even when an attacker cannot see gradients, logits, or token-level feedback. Their method finds visual delay anchors through response-length-guided evolutionary search, then reduces perturbation magnitude through subspace-calibrated decision-based optimization. Across four open-source VLMs, the authors report increased inference latency and energy consumption under a 5% spatial budget, alongside more repetitive outputs. On three commercial VLM , they report longer responses and higher client-observed latency with query budgets of 500 per API. These are reported experimental results, not evidence here of effects across all deployed systems. The authors say the findings motivate generation-aware safeguards; the abstract does not specify which safeguards would work.

Original source

cs.AI · Article

Source details

Original source date: Nov 19, 2026 (2026-11-19).

Date basis: Original submission date in the publisher’s citation record.

Added to Escape Meridian: 2026-11-20T09:01:36.993Z. First observed: 2026-11-20T05:07:49.630Z. Retrieved: 2026-11-20T09:00:52.223Z.

Limits

The supplied evidence is the abstract and page text, not the paper's detailed methods or experimental analysis; it does not identify the four open-source models or three commercial . The reported results concern the tested models and API query budgets; the supplied text does not establish how broadly they generalize or which generation-aware safeguards would work.

Open question

What generation-aware safeguards, if any, would address the final-response-only availability risk described by the authors?

Source-stated

The authors present ChronoPatch as a black-box attack that prolongs VLM generation using only the final response.

We present ChronoPatch, a black-box attack that induces prolonged VLM generation requiring only the final response.

Source-stated

The authors frame prolonged generation as an availability risk because response length affects latency and computational cost.

autoregressive decoding makes response length a key determinant of latency and computational cost.

Source-stated

The method searches for visual delay anchors and then minimizes perturbation magnitude.

It discovers visual delay anchors through response-length-guided evolutionary search, then minimizes perturbation magnitude through subspace-calibrated decision-based optimization.

Source-stated

Across four open-source VLMs, the authors report increased latency and GPU energy consumption under a 5% spatial budget.

Across four open-source VLMs, ChronoPatch increases inference latency by 47.86%--293.34% and GPU energy consumption by 41.44%--249.70% under a 5% spatial budget.

Source-stated

On three commercial VLM APIs, the authors report increased response length and client-observed latency with query budgets of 500 per API.

On three commercial VLM APIs, ChronoPatch increases response length by 37.6%--81.5% and client-observed latency by 1.82x--2.29x under query budgets of 500 per API.

Source-stated

The authors say their findings motivate generation-aware safeguards for deployed generative AI services.

These results reveal a generation-based availability risk that persists under final-response-only access and motivate generation-aware safeguards for deployed generative AI services.

Source and contributor context

Xing, Jiawei, Wang, Li, Meng, Xiangtao, Zang, Chuanchao, Wen, Rui, Li, Zheng · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Latency

The delay between sending a request and receiving the result.

Graphics processing unit

A chip that performs many calculations at once; the main hardware used to train and run AI models.

API

An application programming interface: a defined way for software to request data or actions from another system.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Editorial provenance

Version 1. Publication ID: 06e4f1cd-39dc-44e5-8357-511ef00cbb7b.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-20T09:01:36.993Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 6183064436e0911856ae0068b3865bd4909e2bb7a0ff1be583edbf7e32afca96. Source text hash: 30141bedd111b284396a118959848e5f44fc00b7b1308d52f5475cb8045198cb.

uncertain · 2026-11-20T09:01:36.993Z · anthropic/claude-opus-5.5

Publication: 06e4f1cd-39dc-44e5-8357-511ef00cbb7b

attribution — pass: The reading consistently attributes the attack, its rationale, the results and the safeguard recommendation to the paper's authors. It does not present the reported latency and energy effects as independently verified facts.

We present ChronoPatch, a black-box attack that induces prolonged VLM generation requiring only the final response.

summary — uncertain: The summary never states plainly that the attack works by feeding the model subtly altered images. That is buried in 'visual delay anchors' and 'perturbation magnitude', so a general reader may miss the core mechanism. It also drops the reported magnitudes, such as 47.86%–293.34% latency and 1.82x–2.29x on commercial APIs, which weakens the source's evidence. Smallest repair: add one plain sentence that the attacker supplies adversarially modified images that make the model write longer. Restore the reported ranges. Terms such as black-box, availability risk and spatial budget should get tooltips.

adversarial visual inputs can prolong generation and increase serving overhead

date — pass: originalPublished is 2026-11-19, which matches the arXiv v1 submission date. The feed date is a day later and is correctly not used.

[Submitted on 19 Nov 2026]

support — pass: Every claim matches a verbatim excerpt from the abstract. The caveat about generalization stays within the evidence. The question about which safeguards would work follows from the authors' unspecified recommendation and does not lead the reader.

motivate generation-aware safeguards for deployed generative AI services

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.

Challenge or correct this reading

Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.