Search arXiv⌕ Search

arXiv subjects

Mohammad Zandsalimy

Publications and source records attributed to Mohammad Zandsalimy.

9 recordsLinked to original sources
arXiv2026

Blind, Not Weak: A Best-of-Suite Safety-Utility Frontier for Recover-and-Reguard Defenses Against Encoded VLM Jailbreaks

Safety classifiers ("guards") are the dominant black-box defense for vision-language models, yet a guard judges an input's surface form, not its meaning: a harmful request re-encoded as set theory, formal logic, a classical language, code, or text rendered inside an image slips past a guard that would block it in plain language - the decode gap. The standard fix is a preprocessor that recovers image content and decodes the encoding before the guard. We build one and evaluate it against an ensemble of eleven encoding attacks - six published implementations, one standard encoding baseline, one adapted and three author-constructed renders - counting a behavior as broken if any attack succeeds. Restoring a view the guard never had is what buys coverage - block rates on image renders go from exactly zero to 67-90% - and what it costs in benign traffic is set by the guard, not by the mechanism: one guard pays 9 benign blocking points for the same 70-point gain another pays 69 for. It still does not make the system safer: against an attacker free to choose among eleven encodings, closing one channel relocates the success rather than removing it, and no ensemble contrast for that step survives multiple-comparison correction. What does lower ensemble attack success is a reguard step that re-screens the recovered pre-decode surface, and it is the one every guard pays for: it raises benign over-refusal on all ten guard-target pairs, where restoring a single channel raises it on some and not others. Across the full guard x target x condition factorial, no configuration reaches an ensemble attack-success rate at or below 40% while holding benign over-refusal under 70%. That empty region is a property of the configurations we sample, not a bound on what recovery-based defenses can reach, and we breach its safety half ourselves.

cs.CR↗
arXiv2026

Depth, Not Breadth: Best-of-N Jailbreaking Beyond Surface Noise

Best-of-N jailbreaking spends a query budget on surface variation, scrambling and recasing a request until one draw lands. We ask what a budget buys when its variance is moved into a structural channel instead, holding the search identical across both arms so the encoding is the only difference. Against SAGE, the strongest published self-check defense, best-of-N over a code-completion encoding reaches 67, 22 and 15% of behaviors on three open-weight targets, where that encoding fired once reaches at most 4.7% and the published character search at full budget at most 3.0%: 9 to 75 times the sum of the parts, with bootstrap intervals clearing both ingredients on every target. We report the operative figure beside the headline rather than the headline alone: at the actionable severity threshold those cells read 24, 8 and 1 behaviors (95% CI [13, 28], [3, 13], [0, 3]). A 2x2 holding encoding and variation apart shows the two defense families fail to different factors: a transform defense is broken by the depth of the encoding (7 -> 67 behaviors at fixed variation) and a gate by the breadth of the variation (13 -> 57 at fixed encoding). Repeated sampling also inflates apparent robustness, because an attacker who may try N times experiences the maximum over draws while safety results are reported as means: on one target SAGE blocks 99.8% of individual draws yet loses 12 behaviors to a repeat attacker where a classifier gate blocking 95.6% loses 10. Removing the target's sampling costs SAGE 59, 76, 82 and 29 points of coverage more than it costs an undefended control, against 25, -5, 8 and 2 for a defense whose verdict comes from a fixed shadow model. The design that loses is the one fusing screening and answering into a single generation, so every attacker draw redraws the safety decision as well. The prescription is architectural, not free: do not fuse screening with generation.

cs.CR↗
arXiv2026

The Uncontrolled Variable: Vision-Language Refusal Is Conditioned on the Image-Attachment Interface, and Not Robust to Irrelevant Image Properties

We show that aligned vision-language models also condition refusal on a property of a request's form: whether an image is attached, holding everything the request asks fixed. Attaching a blank canvas, an image that cannot be read, cannot relate to the request, and is byte-identical across every prompt in its condition, shifts benign refusal by tens of points. The shift is not blanket caution but a threshold shift: genuinely neutral instructions are almost unaffected (<=2 percentage points on three of four hosted models) while borderline-benign prompts move +23 to +51 points, so the cost falls on sensitivity-adjacent traffic, meaning benign questions about privacy, self-harm, violence and illegal activity. There is a benign reading of such a threshold, namely that attachment correlates with risk in real traffic, and we take it seriously; a black-box study cannot measure that correlation and we do not claim to. What it can test is whether the response to attachment is robust to variation carrying no information about the request, and on four independent measurements it is not. It varies with canvas colour and pixel count. Its sign inverts across checkpoints. It survives an explicit instruction to disregard the image. And on one model it fires on a bare assertion that an attachment exists, with nothing attached and the modality word contributing none of it. Finally we price it. On a matched harmful set the same canvas does lower attack success, so the cue buys something. But the charge is decoupled from the purchase: the checkpoint with the least harmful headroom we measure, completing only 2% of plain harmful requests, still pays the benign cost in full, and across our models the harmful-side denominator falls as alignment improves while the benign cost does not track it down. Image presence is not a conservative default that a deployer chose and priced. It is an uncontrolled variable.

cs.CR↗
arXiv2026

0%, 45%, or 99%: A Guardrail's Own Share of the Refusals It Is Credited With

A defended pipeline's refusals have two producers: the guardrail bolted in front of the model, and the model's own alignment. Recovering the split costs nothing, because a guard block replaces the model's response and the two counts are therefore disjoint. Holding the defense, the targets, the corpus and the judge fixed, the guardrail's own share of the refusals credited to it is 0%, 41-45%, or 99% across three settings that a results table would describe identically. Two choices move it, and neither belongs to the deployer who bought the guardrail. The attacker drives the share to zero by choosing which channel carries the payload: a plainly written request rendered as pixels, with nothing obfuscated, leaves a text guard's read covering none of it. The evaluator drives the share to 99% by choosing what text fills a defense's internal slots: fill them with the unencoded request behind an encoded attack, a read no deployed defender possesses, and the same guard blocks almost everything. The two consequences differ, and only the attacker's can happen to a running system. The evaluator's choice is an artifact carried by the literature, and its size is set by where the granted text lands: substantial at a guard gate, smaller in a caption-mediated re-check, absent in a majority-vote smoother, an ordering reproduced in an independent replicate. Isolating the grant inside the caption-mediated defense refutes the prediction we registered, since the harm-verdict stage contributes nothing while the stage that regenerates the answer carries the whole effect. The reference implementation builds every stage from a single prompt field that cannot represent the difference between what the attacker sent and what the benchmark records, and an audit of four further released harnesses and of the benchmark itself finds the same structural gap, so faithful porting supplies the grant silently.

cs.CR↗
arXiv2026

Refusing Everything Looks Safe: Restoring the Benign Arm to Encoded-Prompt Evaluation

Encoded-prompt attacks are evaluated almost entirely on their harmful arm: a benchmark sends obfuscated harmful requests and reports how often the model complied. A high refusal rate there is reported as safety, and it is equally consistent with a model that has stopped telling the request apart from anything else in the same format. We run the benign arm through the same transformation, and the two cases are far apart. Across four 7-8B models spanning three base families and four post-training recipes, refusal of harmful homoglyph-encoded prompts spans 0.08 while the same four span 0.57 on the identical requests in plaintext. What the encoding destroys is not refusal but the harm gap: on one model the gap between harmful and benign refusal falls from +0.82 in plaintext to exactly 0.00 under the encoding, and a benchmark reading only the harmful arm scores that model and one retaining a +0.61 gap identically. Running the cell such benchmarks leave out (plaintext content wearing the attack template, with nothing obfuscated) shows that on two of the four models the loss is caused by the protocol rather than by the character transformation, and on a third by the characters. Across a full SFT -> DPO -> RLVR pipeline the harm gap rises by +0.26 with a paired interval excluding zero while the standard harmful-arm metric registers no resolved change at all. We report twelve instrument defects, each with the control that caught it, including a binary jailbreak judge that fires on 0.61-0.70 of responses to plaintext benign prompts; six of the twelve inflate apparent safety, which is the direction a broken safety evaluation fails in by default.

cs.CR↗
arXiv2026

Unread or Unenforced? Separating Representation from Enforcement Failure in Content Guards

When an encoded attack passes a content guard, the guard either never represented the payload's harmful content or represented it and failed to act. End-to-end attack success rate reports one number for both, yet the two have opposite remedies: one is a representational limit that more safety training cannot reach, the other is a decision rule that it can. We separate them by reading a guard's own residual stream, using a content probe fitted on plaintext and transferred without refitting to the encoded condition, alongside the verdict logits from the same pass. Licensing that read honestly is most of the problem and is our main contribution. A conventional permutation test admits the decode measurement on most of a 19-condition encoding ladder for each of two open guards. A length-matched null and a floor calibrated on conditions the guard's base model provably cannot decode reduce it to four conditions each; holding out the items the probe was fitted on removes one more. A third screen constrains the block axis, which the decode screens leave untouched, by running plaintext content inside each condition's own wrapper. It removes the largest cell that survived them. What remains is a policy failure that survives an item-level holdout on two of the four surviving conditions, at 8 and 7 per 100 prompts, against 17 and 23 when the probe is allowed to have seen the prompt it is scoring. It is also confined to one family of surface encodings: where an encoding leaves content linearly recoverable we can separate the two failures, and on genuine ciphers we report the cells as unmeasured rather than as evidence that nothing was decoded. Across every guard and condition pair, blocked without decoding is near zero, so we find little evidence for a pure encoding-format detector under the conditions we test. That cell is the one read we do not repeat under the holdout, and we report it as such.

cs.CR↗
arXiv2026

Decoy Images Amplify Caption-Mediated Defenses Against Encoded Jailbreaks

We report a counter-intuitive interaction between image inputs and existing black-box defenses on Vision--Language Models (VLMs): pairing an encoded jailbreak prompt with an unrelated decoy image can sharply lower attack success rate (ASR). The operative change is in the defense pipeline, not in the image. Across five frontier VLMs, two encoded-attack families, and three black-box defenses, a caption-mediated defense (ECSO) that leaves ASR essentially unchanged on text-only encoded input drops it by up to $73$pp once a content-free decoy is attached; every non-saturated contrast is significant under exact McNemar tests. We advance two hypotheses for this pattern, supported by indirect evidence rather than pipeline introspection, since a black-box threat model precludes inspecting vendor internals: caption-mediated defenses branch on image presence, and intrinsic image-side safety engages on image-resident content. Three controls constrain the explanation. Blank-canvas and natural-photograph decoys reproduce the effect on every model, implicating image presence rather than content; the effect replicates on three open-weight VLMs served with no moderation layer, so it is not a vendor-filtering artifact; and a non-symbolic, meaning-based encoder reproduces it, so it is not specific to symbolic obfuscation. Attaching a decoy unconditionally is not deployable --- it raises benign refusal to $20$--$79\%$, an inflation of $+10$ to $+67$pp --- but gating attachment on a lightweight encoded-input detector returns benign refusal to the text baseline while preserving the safety gain wherever the detector fires, making detector recall the binding constraint. Under adaptive attacks that target the caption-mediated re-check, the effect degrades but holds. We frame this as an observation about pipeline interaction, not as a robust defense.

cs.CR↗
arXiv2026

Exposing LLM Safety Gaps Through Mathematical Encoding:New Attacks and Systematic Analysis

Large language models (LLMs) employ safety mechanisms to prevent harmful outputs, yet these defenses primarily rely on semantic pattern matching. We show that encoding harmful prompts as coherent mathematical problems -- using formalisms such as set theory, formal logic, and quantum mechanics -- bypasses these filters at high rates, achieving 46%--56% average attack success across eight target models and two established benchmarks. Crucially, the effectiveness depends not on mathematical notation itself, but on whether a helper LLM deeply reformulates the harmful content into a genuine mathematical problem: rule-based encodings that apply mathematical formatting without such reformulation perform no better than unencoded baselines. We introduce a novel Formal Logic encoding that achieves attack success comparable to Set Theory, demonstrating that this vulnerability generalizes across mathematical formalisms. Additional experiments with repeat post-processing confirm that these attacks are robust to simple prompt augmentation. Notably, newer models (GPT-5, GPT-5-Mini) show substantially greater robustness than older models, though they remain vulnerable. Our findings highlight fundamental gaps in current safety frameworks and motivate defenses that reason about mathematical structure rather than surface-level semantics.

cs.CR↗
arXiv2025

Behind the Mask: Benchmarking Camouflaged Jailbreaks in Large Language Models

Large Language Models (LLMs) are increasingly vulnerable to a sophisticated form of adversarial prompting known as camouflaged jailbreaking. This method embeds malicious intent within seemingly benign language to evade existing safety mechanisms. Unlike overt attacks, these subtle prompts exploit contextual ambiguity and the flexible nature of language, posing significant challenges to current defense systems. This paper investigates the construction and impact of camouflaged jailbreak prompts, emphasizing their deceptive characteristics and the limitations of traditional keyword-based detection methods. We introduce a novel benchmark dataset, Camouflaged Jailbreak Prompts, containing 500 curated examples (400 harmful and 100 benign prompts) designed to rigorously stress-test LLM safety protocols. In addition, we propose a multi-faceted evaluation framework that measures harmfulness across seven dimensions: Safety Awareness, Technical Feasibility, Implementation Safeguards, Harmful Potential, Educational Value, Content Quality, and Compliance Score. Our findings reveal a stark contrast in LLM behavior: while models demonstrate high safety and content quality with benign inputs, they exhibit a significant decline in performance and safety when confronted with camouflaged jailbreak attempts. This disparity underscores a pervasive vulnerability, highlighting the urgent need for more nuanced and adaptive security strategies to ensure the responsible and robust deployment of LLMs in real-world applications.

cs.CR↗