A corrected protocol and a validated leak diagnostic are presented, not a new OOD method: under the corrected protocol, perturbation signals are decodable but not detectable, and the perturbation method does not improve on plain Mahalanobis distance.
Abstract
While auditing a perturbation-based OOD detector on a document benchmark, we recorded an AUROC of 0.326 -- well below the 0.5 chance level. The cause is a benchmark leak: the designated"OOD"class is one the model was trained on, so its examples sit inside the in-distribution fit set and the detector is penalized for correctly ranking them as familiar. Deleting the class and retraining 35 models across two domains raises the score to 0.911. We distill the contamination into a leak fingerprint -- near-perfect supervised decodability (AUROC approximately 1) coupled with unsupervised detection collapsed below 0.65 -- and validate it on a controlled battery of 52 settings (20 leaked, 32 clean) across ResNet-50 and ViT-B/16 on CIFAR-10/100, achieving sensitivity 18/20 and specificity 31/32 in embedding space; the matched fit-set-exclusion controls are perfect at 20/20. An in-the-wild audit of 24 standard near/far OOD benchmark pairs fires on exactly one (the intrinsically hard CIFAR-100 vs CIFAR-10 pair) and on no far-OOD pair, confirming specificity and that standard cross-dataset construction is clean. Under the corrected protocol, perturbation signals are decodable but not detectable: a supervised reader recovers the OOD signal (AUROC 0.87-1.00) while no unsupervised detector does, and the perturbation method does not improve on plain Mahalanobis distance. We provide a theoretical account of why and, for transparency, retract an earlier circular correlation. The contributions are a corrected protocol and a validated leak diagnostic, not a new OOD method.
Context can change whether a request is harmful without changing its topic or surface form. We ask whether residual-stream probes distinguish harmful requests from surface-matched benign controls at a useful operating point. Across three 7-8B model families, an activation sensor blocks 95.5-97.7 percent of judge-classified compliant attacks in a taxonomy-selected set. It also blocks 59.6-68.4 percent of XSTest prompts. A fully disjoint audit reconstructs near-ceiling source-contrast AUROC (0.996-0.999), but fixed transfer to matched pairs is weaker: 0.656-0.819 on the guard-selected Twin-n70 subset and 0.590-0.690 on the full Twin-n163 cohort. We test ten axes on the reference family and seven across all families with leakage, hold-out, and permutation controls. On Twin-n163, no axis evaluated without direct pair-boundary fitting reaches the specified numerical threshold. Requiring persistence on that full cohort was added at analysis time. A separately specified 24B/32B extension gives the same result. Pair-trained classifiers weaken under category and generation-batch hold-out and false-block 79.6-100 percent of XSTest at 95 percent in-corpus TPR. At the tested read points, these activation scores behave as broad-risk detectors rather than standalone context adjudicators.
Benchmarks for systems that are optimized against the evaluation signal measure something different from what they claim. We document this concretely in two GPU-kernel-optimization suites with held-out generalization gates: Metal-Sci (10 scientific-compute tasks) and Metal-ZK (12 zero-knowledge/cryptographic tasks), in which three frontier LLMs (Opus 4.7, Gemini 3.1 Pro, GPT-5.5) propose Metal kernels inside a $(1{+}1)$ evolutionary loop with rich feedback. Although no model is prompted to act adversarially, the promoted winners repeatedly fingerprint the evaluation configuration: they branch on the identity of runtime parameters, tune the measured branch maximally, and leave the unmeasured branch slow or silently wrong. Across the pooled suites, $16/53$ ($30\%$) of in-distribution wins fail to transfer to held-out configurations. We give a four-mode taxonomy of these failures, from configuration fingerprints to gate leakage. We distill design guidance for measurement under strategic optimization: held-out probes retain validity only on non-enumerable axes; gates must measure held-out performance, not just correctness; and a transfer rate is interpretable only with per-failure mechanism grades: ours decomposes into gamed, overfit, and benign. Code and research artifacts: https://github.com/vicgalle/kernel-fingerprinting
Almost all adversarial attacks add an imperceptible perturbation to fool a model. We instead study the opposite: a large, clearly visible perturbation that causes the model to keep its original, correct prediction, even though a human would no longer recognize the image. Prior work showed such examples can be generated at scale but left three questions untested: whether humans really perform worse than the model, whether standard out-of-distribution (OOD) detection and calibration tools catch it, and whether existing defenses mitigate it. We answer all three on MNIST, CIFAR-10, and ImageNet. (i) An independent recognizer proxy drops to ~49% on CIFAR-10 while the model stays at 100% -- a gap a small human pilot (N=5) corroborates directly and that is not explained by signal loss (a matched-magnitude Gaussian control degrades recognizability faster); a CLIP zero-shot proxy confirms the gap at ImageNet scale too. (ii) Confidence- and energy-based OOD detectors and calibration are structurally blind (0% detection, ECE ~= 0), while a feature-space Mahalanobis detector flags 100% -- but is evaded by an adaptive attacker at no cost to success. (iii) No classical defense, including adversarial training (45% robust accuracy), reduces attack success (correlation with large-epsilon_l resistance r ~= 0). A mechanistic analysis further shows the attack destroys low-level texture far faster than edge/shape structure.
Knowledge distillation enables an adversary to replicate a proprietary classifier by querying its prediction interface and training a surrogate on the returned probability vectors. Antidistillation sampling, proposed for large language models, counters this threat with an input-dependent, gradient-directed perturbation of the served distribution; its transfer to classification has not been studied. Adapting the defense to classification, we show its behavior is governed by the distribution of the teacher's per-input confidence margins. Because well-trained classifiers are severely overconfident, the direct transfer exhibits an inert window: below a closed-form-predictable threshold, it affects neither attacker nor defender; beyond it, the defense undergoes a phase transition and degrades the teacher faster than the attacker's student. Temperature softening rescales the transition in closed form, and every temperature configuration lies on the same unfavorable trade-off curve. Our method, ADS-C, composes the perturbation under a closed-form, per-input margin budget that provably preserves every served top-1 prediction, so the defended teacher's accuracy equals the undefended teacher's identically. Under this guarantee the distilled student still loses 17.4 percentage points on CIFAR-100, 29.6 on CIFAR-10, and 13.3 on Tiny-ImageNet; matching this degradation with the unmodified defense costs 27.5, 32.9, and 22.2 points of teacher accuracy. Because served labels are unchanged, a hard-label attacker gains nothing, while the defended soft output trains a student up to 29.7 points below that floor: the incentive to distill served probabilities is not merely removed but reversed. To our knowledge, ADS-C is the first antidistillation defense for classification whose utility cost is exactly zero.
For AI agents to be useful beyond simple chat, they must hold sensitive user context such as calendars, credentials, health records, and financial data. We study whether the mere presence of such secrets in a model's context window introduces hidden correlations into the model's benign outputs, allowing reconstruction even when the model correctly refuses direct extraction. We further study whether an adversary can actively engineer prompts that amplify this effect, using the model as a covert carrier to transmit secrets through seemingly innocuous text. In both cases, this limited leakage is exploited using a novel adaptive attack that assumes black-box access to the underlying model. In controlled experiments across eight proprietary models, we find that 2-digit in-context secrets are reconstructed with near-perfect accuracy and 4-digit secrets at 82\% exact match, all from outputs the model produces in response to ordinary, non-adversarial requests. We observe that more capable models leak more: stronger instruction-following amplifies sensitivity to in-context secrets, suggesting leakage is a byproduct of capability as opposed to a patchable bug. We show this leakage enables two practical attacks: (1) a trained classifier that infers semantic predicates about user memories (e.g., health conditions, financial events) from routine natural-language outputs, and (2) an RL-trained adversary that extracts full Social Security Numbers from a production-style agent.
Jaiden Fairoze, Neal Mangaokar, Kamalika Chaudhuri et al.· 0 citations
Large language models (LLMs) are increasingly consumed through opaque serving chains - API aggregators, resellers, and inference providers - in which the client has no technical means to confirm that the model answering is the model advertised, and recent audits show that a substantial fraction of commercial endpoints deviate from the vendor's reference weights. Existing identification techniques require long generated texts, token-level log-probabilities, adversarially crafted prompts, or the model owner's cooperation. We show that far weaker evidence suffices. We define a behavioral fingerprint of an LLM as the empirical distribution of its answers to trivial one-word prompts -"name a random number between 1 and 100"- collected across four languages at a cost of one output token per query. Measuring 165 models served via a large commercial aggregator (OpenRouter), we find that (i) these distributions are highly non-uniform (median cell entropy 1.0 bit) and model-specific: split halves of the same model's samples lie an order of magnitude closer than samples of different models; (ii) Jensen-Shannon divergence between fingerprints recovers model lineage, assigning a model to its documented family with 59.5% leave-one-out accuracy against an 18.4% chance rate; and (iii) a biometric-style verification protocol achieves a 7.3% equal error rate with the full 40-cell battery, and below 11% with eight probe cells - roughly a hundred single-token queries per audit. We further report ecosystem anomalies, including a proprietary-branded flagship endpoint distributionally indistinguishable from an open-weight Qwen model. The protocol, prompts, raw data, and analysis code are released for reproduction and operational use.
