arXiv subjects

Xiangchen Guan

Publications and source records attributed to Xiangchen Guan.

8 recordsLinked to original sources

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

Trust Is Not Enough: Influence Calibration for On-Policy Self-Distillation in Agentic RL

On-policy self-distillation (OPSD) gives language agents dense token-level supervision from a privileged self-teacher on the policy's own trajectories. Existing methods allocate this supervision mainly by teacher trust, but trust does not reveal whether emphasizing a token supports the current policy objective. We call this the trust-utility mismatch and introduce Influence Calibration for Self-Distillation (ICSD). For each supervised token, ICSD measures the first-order response of its importance-weighted RL surrogate contribution to a teacher-directed output perturbation. Batch-adaptive calibration converts this non-stationary signal into a bounded allocation weight while preserving the original auxiliary-loss mass within each action turn. These detached weights affect only the distillation loss and require no additional model pass. Across ALFWorld, WebShop, and Search-QA, ICSD improves all matched aggregate metrics over trust-only allocation under Group Relative Policy Optimization (GRPO) and Group-in-Group Policy Optimization (GiGPO), across two model families spanning 1.5B to 7B. At 7B, it reaches 96.1% ALFWorld success and a WebShop score of 93.1. Frozen-batch analyses show that ICSD reduces teacher-supported mass assigned to objective-opposed tokens from 60.1% to 37.8% and raises cosine compatibility with the RL gradient by 0.192. A companion repository is avail- able at https://github.com/lanqz7766/Influence-Calibration-for-On-Policy-Self-Distillation-in-Agentic-RL.

cs.AI↗

arXiv2026

The Uncontrolled Variable: Vision-Language Model Refusal Responds to Image Presence in Ways Risk Cannot Explain

Vision-Language Model (VLM) safety is expected to depend on what a request asks for. We show that safety-aligned VLMs also key refusal on a property of a request's form: whether an image is attached, holding everything the request asks fixed. Attaching a blank canvas - unreadable, unrelated to the request, identical across prompts - shifts refusal by tens of percentage points, with no defense in the loop. The shift is not blanket caution. Neutral instructions are almost unaffected while borderline-benign prompts move sharply, so the cost falls on sensitivity-adjacent traffic: benign questions about privacy, self-harm and violence. Attachment alone is sufficient, while the image's properties set the price: a black canvas costs substantially more than a white one of identical size, and on an open checkpoint the carrying axis is pixel count. Nor is the shift under instructional control - telling the model the image is a placeholder to be disregarded removes only a fraction of it, and on one model asserting that an attachment exists moves refusal substantially with nothing attached. Attachment may correlate with risk in deployment; what these models do with it does not track risk. It is not the serving stack, since the same weights reached two ways behave alike, nor a property of VLMs as such, since several open-weight checkpoints show nothing. It belongs to particular aligned checkpoints, one of them open. It is also decoupled from what it buys: the canvas does prevent some attack success on a matched harmful set, but far less than it costs, and its sign is not fixed - on one open model the identical canvas makes the model markedly easier to attack. Image presence is not a default a deployer chose or priced; it is an uncontrolled variable inherited with the weights.

cs.CR↗

arXiv2026

Attack Ensembles Expose a Safety-Utility Trade-off in Black-Box Guard 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 published encoding attacks, counting a behavior as broken if any attack succeeds. That metric separates two mechanisms such defenses conflate. Restoring a view the guard never had improves it on both axes at once: it blocks far more attacks, and, measured on a category-balanced benign set, it blocks fewer benign requests, because restating a request normalizes the borderline phrasing a classifier over-flags. 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 survives multiple-comparison correction. What does lower ensemble attack success is re-screening the recovered pre-decode surface, and that step is where the entire benign cost falls. The safety-utility trade-off is therefore not a property of recovery; it is localized to one step. 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%. The per-attack averages usually reported understate the attacker roughly fourfold, which is why this frontier is easy to miss. Composing across defense families is the one lever that moved the safety axis, beating every configuration we measured, and still landing far outside any deployable refusal budget.

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

MemPrism: Task-Conditioned Relational Memory Views for Long-Horizon Agents

Long-horizon agents rely on memory to reuse experiences, yet existing memory systems often assume that evidence can be directly consumed through a fixed representation. This leads to representation mismatch, where relevant information is available but not organized for the current decision. To this end, we propose MemPrism, a task-conditioned relational memory framework that separates persistent experience storage from decision-time working memory. MemPrism records interactions as the event stream and dynamically constructs relational views according to the current task context. A lightweight view policy selects the relation structure, evidence range, outcome condition, and granularity, while a deterministic composer and render transform historical facts into a temporary optical working-memory view for a frozen task policy. Experiments on long-horizon embodied and web-agent benchmarks show that MemPrism consistently improves the task performance, especially as trajectories become longer, while reducing memory token consumption. Furthermore, the learned view policy transfers across different VLMs without additional adaptation, demonstrating the effectiveness of task-conditioned relational views as a general memory interface for agents.

cs.AI↗

arXiv2026

Borrowed Strength: Best-of-N Search over a Code EncodingBreaks Self-Check Jailbreak Defenses

A self-check defense asks the target model to assess a request before answering it; SAGE, the strongest published instance, reports an average 99% defense success rate. We show it can be breached by composing two attacks that are individually harmless against it: an established code-completion encoding and an established best-of-N search, neither of which exceeds 4.7% of behaviors alone. Composed, with the search budget spent on the encoding, they reach 67/22/15% across three open targets, and the effect persists on a 70B target. We then explain the composition rather than only reporting it. First, a self-check defense borrows its strength from the target: SAGE does not detect the attack, it asks the model to, and the four targets convert that request into an explicit refusal between 32% and 97% of the time, which orders the spread in defended coverage even though undefended reach is near-identical. Second, which attack survives is decided by the type of defense, and it inverts: against transform defenses the code encoding retains far more of its undefended reach than the character search, while against gate defenses the ordering flips. We account for this with the number of independent probes an attack delivers to a defense's decision boundary. Finally, we report a validity defect we found and repaired in our own pipeline, a deterministic attack under greedy decoding has no best-of-N variation channel at all, and give the one-line diagnostic that detects it. All claims rest on 310,000 generations scored by a human-validated judge.

cs.CR↗