A drug called memantine is remarkably safe for a reason that has nothing to do with receptors in the body: it only blocks a channel while that channel is stuck open in a pathological state, and slips off again during normal, brief signaling. This proposal imports that principle — uncompetitive, open-state antagonism — into adversarial machine learning. Today, every adversarial perturbation is static: once baked into pixels or tokens, it acts continuously, whether or not the target model is in a vulnerable state. We propose state-gated adversarial perturbations that are measurably dormant during ordinary inference and only begin to suppress a model's safety machinery when that machinery starts to fire. The practical payoff is stealth of a new kind: an attack that leaves clean behavior untouched, is invisible to monitors watching for abnormal safety-circuit activity during benign traffic, and is cheaper than existing circuit-level attacks.
The biological side of this proposal is a design blueprint, not an application domain. Selective receptor antagonism is a mature discipline built on one insight: blunt, always-on blockers of a receptor family poison the homeostatic functions alongside the pathological ones, so useful drugs must discriminate — by subunit, by spatial compartment, or by activity state (Zheng et al.). Machine learning safety is arriving at the same fork. Sun et al. showed that memantine spares synaptic signaling while selectively reversing the extrasynaptic cascade that kills neurons, precisely because it gates on channel state; broad blockers like MK-801 instead cause widespread dissociative toxicity (Wang et al.). The computational analogue is already visible: Xie et al. found that safety refusal in vision-language models concentrates in a sparse internal "defensive circuit" nearly disjoint from the semantic circuit (Jaccard index under 3%), and Pramanik et al. showed refusal is driven by a dynamic subset of attention heads. What no one has built is the memantine of adversarial attacks: an intervention whose potency is conditional on the very state it suppresses.
Core Claim
An adversarial perturbation can be optimized so that its effect on a model's internal representations is negligible whenever the model's defensive circuit is at rest, and strongly suppressive only when that circuit is active. Formally, on a white-box surrogate we extract a linear probe direction $d$ for defensive-circuit activation (following the probing recipe of Xie et al.), define the defensive activation $a(x) = d^\top h(x)$ where $h(x)$ is the model's representation at the defensive bottleneck, and optimize a perturbation $\delta$ against:
$$\mathcal{L}(\delta) = \mathbb{E}_{x \in \mathcal{B}} \|\Delta h(\delta, x)\|^2 \;+\; \lambda\, \mathbb{E}_{x \in \mathcal{H}} \big( g(x)\, s - \Delta a(\delta, x) \big)^2$$
Here $\mathcal{B}$ is benign traffic (e.g., VQAv2 questions) and $\mathcal{H}$ is harmful prompts (e.g., AdvBench); $g(x) = \sigma(a_{\text{clean}}(x) - \theta)$ is a gate that is near zero when the unperturbed model is at rest and near one when it is beginning to refuse; $s$ is a target suppression amount; $\Delta h$ and $\Delta a$ are the perturbation-induced shifts in representation and defensive activation. The first term enforces dormancy; the second enforces open-state blockade. A naturalness constraint (spectral smoothness as in Xie et al., or diffusion-manifold projection as in Yao et al.) keeps the trigger perceptually innocuous. The mechanism is not magic: the gate works because the model's own attention to the perturbed tokens changes as it enters a refusal state, and the objective trains the perturbation to exploit exactly that interaction.
This differs from the closest existing work in one specific way. SCIA (Xie et al.) selectively suppresses the defensive circuit but does so unconditionally — the perturbation fights the safety circuit on every input, requires roughly 50 hours of A100 compute per AdvBench run, and would be flagged by the circuit-aware defenses its own authors anticipate. HMNS (Pramanik et al.) achieves an allosteric-like effect by steering in the nullspace of refusal heads, but only through direct white-box weight manipulation at inference time. PSI (Yao et al.) constrains where the perturbation lives (the natural-image manifold) but not when it acts. Our claim is about the when.
Research Program
Three components. First, defensive-state cartography: on open surrogates (LLaVA-1.5 and one LLM from the HMNS model set), replicate the extraction of defensive-circuit probes and, extending prior work, measure how the probe's activation trajectory evolves across generation steps — Pramanik et al. showed refusal heads are a moving target, so the gate $g$ must be defined over a trajectory, not a single forward pass. Second, gated optimization: train $\delta$ with the objective above, including a per-step variant where the gate reads the model's own intermediate state rather than the input's harmfulness label — this is what makes the mechanism genuinely state-gated rather than input-class-gated, and it is the ablation that isolates our key ingredient. Third, dormancy measurement: we introduce a collateral-shift metric, the KL divergence between the model's outputs with and without $\delta$ on benign traffic, alongside clean-utility benchmarks, because attack success rates alone cannot distinguish a memantine from an MK-801.
We will also run the proposal's central defensive stress test. Because a state-gated trigger need not be spatially localized, we evaluate against GuardAlign, the strongest selective test-time defense in the corpus, which masks individual unsafe image patches via optimal transport. GuardAlign's authors explicitly leave open whether a globally entangled perturbation can defeat patch-level masking; our diffuse, state-gated trigger is a concrete candidate, and a companion ablation with the same trigger confined to a localized patch (à la Guo et al.) tests whether diffusion of the trigger across the image is what breaks the defense.
Baselines
The naive reference point is the undefended model doing nothing: vanilla LLaVA-1.5 has an unsafe response rate of 46.04% on the SPA-VL benchmark and 78.51% accuracy on VQAv2, both reported by GuardAlign. On a defended model, any attack must first be measured against the defense itself: GuardAlign cuts the unsafe rate to 10.31% while raising VQAv2 accuracy slightly to 79.21%, so the bar for a successful attack is pushing the defended model's unsafe rate materially above 10.31% while holding its VQAv2 accuracy within about a point of 79.21%.
The strongest existing competitor is SCIA (Xie et al.), which reports an average attack success rate of 59.7% across open-source models on AdvBench and 46.8% against GPT-4o, at a cost of approximately 50 A100-hours for the full AdvBench set. We will re-run SCIA on our identical model and benchmark suite under a matched GPU-hour and tuning budget, and compare on four axes at once: ASR on AdvBench and SPA-VL, collateral shift on benign traffic, perceptual and judge-based stealth, and wall-clock compute. Beating SCIA means matching its ASR within a few points while cutting collateral shift by an order of magnitude and reducing compute by at least a factor of five; if SCIA matches us on ASR but we win decisively on dormancy and compute, the tradeoff result still stands on its own.
PSI (Yao et al.) is the strongest stealth-and-transfer baseline, with a 78.6% success rate and a stealthy success rate of 62.8% against GPT-5; we compare against its reported numbers on the commercial gauntlet rather than re-running it, since its pipeline is diffusion-based and its budget is not directly commensurable. Finally, the key ablation is our own optimization with the gate term removed — an unconditional circuit-targeted perturbation with identical naturalness constraints and compute. This isolates state gating as the ingredient, rather than circuit targeting or smoothness, whenever it wins.
Why not just…?
Why not just suppress the safety circuit directly, as SCIA and HMNS already do? On raw attack success, this is a real threat: SCIA's 59.7% average ASR is strong, and nothing in our mechanism should beat unconditional suppression at maximal aggression. The honest answer is that we concede raw ASR parity and claim a different frontier. Static suppression acts on every input, so it measurably degrades benign safety behavior and is exactly what the circuit-aware monitoring that Xie et al. anticipate would catch. Our decisive experiment is a monitoring defense that flags inputs whose defensive-circuit activations deviate from clean statistics: we predict the static ablation triggers it on benign traffic while the gated trigger does not, and if that prediction fails, the proposal's core value collapses to a compute saving.
Why not just use a diffusion-manifold attack like PSI? PSI's transfer numbers (78.6% on GPT-5) are the best in the corpus, so if state gating destroys transferability, our idea loses its practical audience. But PSI's perturbation is semantically present at all times — its own authors note failures on low-entropy backgrounds and detection by scene-level reasoning — and it does not engage the question of dynamic safety circuits at all. The settling experiment is a transfer subset evaluated against commercial models with an LLM-as-judge stealth filter, reported as a stealthy success rate directly comparable to PSI's 62.8% on GPT-5; we treat falling far short of that number on the stealth metric as a failure of the proposal, not a reframe.
Why not just do black-box prompt search or mirror distillation? Iterative searchers like PAIR are ruled out by the evidence: ShadowBreak reports that PAIR needs 27.4 detectable jailbreak queries per sample, which rate limiters and moderators catch. ShadowBreak itself is the genuine threat here: 80% ASR against GPT-3.5 Turbo with only 1.5 detectable queries per sample, at trivial cost ($3.20 in API fees and $8.00 in GPU time). We do not claim to beat it on cost. The difference is deploy-time behavior: ShadowBreak's transferred prompts are explicit jailbreak text, inspectable at the moment of use, whereas our trigger is inert content whose benign-traffic outputs are statistically indistinguishable from clean inputs. If a content-inspection experiment shows our deployed trigger is flagged as often as ShadowBreak's prompts, we are wrong and will say so.
Decisive Checks
Three results decide the proposal. First, on GuardAlign-defended LLaVA-1.5, the gated trigger must raise the SPA-VL unsafe response rate well above the defended 10.31% while keeping VQAv2 accuracy within roughly a point of 79.21% — a memantine-like separation of attack efficacy from collateral damage. Second, the ungated ablation must show either higher benign-traffic collateral shift or detection by the circuit-monitor defense under matched ASR, establishing that the gating term, not circuit targeting generally, does the work. Third, on the commercial gauntlet, the stealthy success rate must be competitive with PSI's reported 62.8% on GPT-5; no reported result found on Lacuna for any state-conditional adversarial perturbation, so any positive number here is the first of its kind. If the first two hold and the third fails, the proposal survives as a defense-evasion and stealth result; if the first fails, the biological analogy was decorative and the idea should be abandoned.
References
- Drugs for Autoimmune Inflammatory Diseases: From Small Molecule Compounds to Anti-TNF Biologics
- Extrasynaptic NMDA receptor-induced tau overexpression mediates neuronal death through suppressing survival signaling ERK phosphorylation
- Differential roles of hippocampal glutamatergic receptors in neuropathic anxiety-like behavior after partial sciatic nerve ligation in rats
- Dissecting the Safety Circuit: Neuronal Intervention for Transferable Adversarial Attacks on VLMs
- Jailbreaking the Matrix: Nullspace Steering for Controlled Model Subversion
- Transferable and Stealthy Adversarial Attacks on Large Vision-Language Models
- GuardAlign: Test-time Safety Alignment in Multimodal Large Language Models
- CBV: Clean-label Backdoor Attacks on Vision Language Models via Diffusion Models
- Stealthy Jailbreak Attacks on Large Language Models via Benign Data Mirroring