Inference-Time Defense

Large language models are often fine-tuned, shared, or downloaded from third parties, so a deployed model may carry a hidden backdoor that behaves normally on benign inputs but switches to attacker-controlled behavior when a secret trigger appears. While backdoors can be audited before deployment, runtime monitoring remains important for models that are frequently updated. The challenge is that LLM serving is latency-sensitive: existing inference-time detectors either rely on assumptions about the trigger form, which can fail on stealthy attacks, or require extra model computation, such as input perturbations or an additional generation pass. We introduce SpecGuard, an inference-time backdoor detector that repurposes speculative decoding at zero added model-computation cost. Speculative decoding speeds up inference by using a small draft model to propose tokens and a target model to verify them. We observe that this verification process already exposes a useful signal: when a backdoor is triggered, the target model shifts toward the attacker's behavior, while a clean draft model does not predict this shift, causing the draft-token acceptance rate to change. We formalize when this signal appears and show that an attacker who suppresses it must also weaken the backdoor. Across diverse backdoor types and model families, SpecGuard reliably detects triggered behavior, including stealthy cases where input-level filters are blind, while avoiding the extra generation cost of existing runtime detectors. Speculative decoding therefore doubles as a free, always-on signal for detecting backdoored LLM behavior.

A prominent paradigm in inference-time alignment employs lightweight supervisors to steer Large Language Models (LLMs). Through empirical analysis, we identify a structural mismatch in this paradigm: weak supervisors exhibit pervasive high entropy across the vast majority of tokens, yet prevailing dense intervention approaches mandate supervision at every decoding step. This leads to frequent low-confidence interventions that can disrupt valid base-model reasoning and incur substantial utility costs. To resolve this, we propose TUSA (Trust-based Uncertainty Sparse Alignment). Moving away from continuous oversight, TUSA reframes alignment as a dynamic arbitration process, introducing an uncertainty-aware arbiter that authorizes intervention only when two conditions are met: the supervisor is confident and the token is semantically salient. This mechanism effectively filters out uncertainty-driven noise and redundant supervision. Extensive experiments across multiple models and benchmarks show that TUSA consistently improves both safety alignment and general helpfulness. By bypassing approximately 50% of alignment steps, it not only enhances safety preference by up to 15.6%, but also boosts general preference rates by up to 12.0% compared to the dense baseline, demonstrating that selective, high-precision alignment can outperform continuous supervision.

Large language models (LLMs) finetuned for specialized domains represent crucial high-impact applications. Inference-time alignment improves safety degraded from specialization finetuning without requiring substantial computational resources, complementing finetuning-based methods with an easy-to-use, plug-and-play solution. However, existing inference-time methods fail to reliably improve safety without disrupting domain capability. We identify the root cause as complementary expertise orthogonality: specialized base models and general-domain guidance models have orthogonal competencies, making the guidance signal unreliable for specialized generation. This primarily manifests as stop token interference, where the guidance model's tendency toward continuation overrides the base model's decision to stop, burying correct answers under guidance-induced continuation. To address this problem, we propose CREST, an inference-time alignment method that steers base model hidden representations using safety directions extracted from a guidance model of any family, avoiding token-level structural limitations entirely. CREST improves safety where specialization has weakened it while preserving both domain-specific capability and the safety of already well-aligned models, outperforming baselines by up to 22.2% on safety benchmarks. Our code is available at: https://github.com/DecayingSeart/CREST.

Large language models (LLMs) frequently generate source code containing vulnerabilities, yet little work studies the internal mechanisms that distinguish safe from vulnerable generation in them. In this work, we systematically perform a mechanistic interpretation of LLMs, aiming at both understanding how code safety-vs-vulnerability is represented or driven by components in an LM and turning the insights into actionable steering strategies to encourage safer code generation. To this end, we introduce CodeSec-Pairs, a dataset of 9,342 Python safe-and-vulnerable contrastive code pairs, sampled from Llama-3.1-8B-Instruct. Utilizing the dataset, we explore approaches to localize layers and attention heads that relate to code safety, and further experiment with different steering strategies for inference-time vulnerability reduction. In particular, we propose DuoSteer, a double-steering approach that simultaneously applies safety and code-correctness steering to attention heads. In experiments over five vulnerability types, DuoSteer leads to an average of -26.9% vulnerability rate reduction and +7.5% functional correctness improvement, which outperforms not only other steering variants but also prompting and supervised fine-tuning baselines. The advantage also replicates on Qwen-2.5-Coder-7B-Instruct with another 2,500 contrastive pairs sampled from that model.

Modern image editors combine vision-language models (VLMs) with diffusion transformer backbones to modify a single reference image according to instructions without fine-tuning. This capability also enables unauthorized manipulation of publicly released images. Existing inference-time defenses either invalidate edits through conspicuous corruption, thereby exposing the protection, or allow them to proceed with identity or reference content drift, thereby failing to prevent the editing behavior itself. We instead target a stealthy and harmless no-op in which the requested edit is suppressed, the output remains natural and source-preserving without conspicuous artifacts or identity replacement, and harmful semantics requested by malicious instructions are absent. We propose NullEdit, which targets the VLM representation jointly formed from the reference image and instruction before it conditions the downstream DiT backbone. Using normal-edit and no-edit anchors, NullEdit redirects this representation, while cross-prompt gradient averaging transfers protection to held out instructions. Across Step1X-Edit and Qwen-Image-Edit on CelebA-HQ and VGGFace2, NullEdit reduces the EditReward IF score by 0.813 on average relative to the SOTA baseline while preserving subject identity and source content.

Ensuring safety and policy compliance in text-to-image diffusion models remains a critical challenge, as benign or adversarial prompts can often elicit prohibited content, e.g. nudity and protected intellectual property. While training-based unlearning methods are effective, they are computationally expensive and prone to catastrophic interference with general capabilities. Conversely, existing test-time defenses are primarily prompt-centric, relying on modifying textual descriptions only, and overlook the visual signals for detection. In this paper, we propose to leverage the intermediate clean image estimated during the generation process and employ a sparse margin objective to detect prohibited concepts. When a violation is detected, we immediately intervene by optimizing a structured low-rank residual in the text-conditioning space via truncated backpropagation. This design allows weight-preserving detection, keeps non-violating inference latency nearly unchanged as the maximum budget increases, and offers flexibility in safety performance via test-time scaling. Extensive experiments on Stable Diffusion v1.4 and v3.5 across nudity removal, IP protection, and style erasure demonstrate superior performance across suppression, fidelity and preservation compared to prior weight-preserving baselines, providing a scalable and flexible solution for safe generative deployment.

AI agents extend large language models (LLMs) with external tools, enabling them to perform complex tasks and translate model outputs into consequential real-world actions. Yet LLMs often become substantially less safe when deployed as agents, and the source of this degradation remains poorly understood. In this paper, we identify schema-formatted tool specifications as a primary source of agent safety degradation and show, through white-box representation analysis, that they weaken the model's internal refusal signals and contribute to unsafe tool execution. Building on this finding, we propose SafeKeep, an inference-time safeguard that decouples safety judgment from tool execution: it assesses requests using flattened textual tool specifications while retaining the original schema-formatted specifications for execution. Across two representative benchmarks and four LLMs, including both white-box and black-box models, SafeKeep increases the average refusal rate for harmful requests from 23.8% to 70.6% and reduces the average attack success rate under observation-level prompt injection from 25.6% to 2.5%. It also outperforms existing safeguards and preserves task-handling capability. We release the code and data at https://github.com/snowcatsmoking/SafeKeep .

Inference time defences against vision language model jailbreaks often subtract a calibrated direction from the residual stream at a chosen decoder layer. We compare five defence candidates across 15 model and layer cells from four architectural families under a magnitude controlled protocol that matches the intervention size for each prompt and pairs every direction with a random control of the same norm. The candidates are the mean image conditioning shift, a CMRM style refusal direction, a ShiftDC style attack specific residual, a prompt instruction to ignore the image, and a random control. No single candidate dominates on both refusal recovery and utility preservation. The image conditioning shift leads on LLaVA 1.5 and Pixtral 12B and is the only candidate whose utility loss remains at the measurement noise floor in every family. The prompt instruction leads on Qwen2.5 VL, while the attack specific residual leads on Qwen2 VL 2B. The image conditioning direction is direction specific in 13 of 15 cells, but strongly architecture specific and nontransferable across the only dimension compatible pair, LLaVA 1.5 13B and Pixtral 12B. We also connect text only and multimodal refusal geometry. The CMRM direction has positive cosine alignment with the image conditioning shift in all 15 cells, with mean 0.35, range 0.17 to 0.65, 15 to 25 times the random vector null, and a sign test p value of about 3e-5. These results show that the two recipes recover partially overlapping geometry and that direction based defences should be calibrated separately for each language decoder family.

This work presents LeakyLMs, a set of attacks that leak proprietary model, architecture, and deployment information from production language models. LeakyLMs is the first to demonstrate that key model and deployment details can be inferred using only token generation timing, even when interacting through remote APIs. LeakyLMs introduces two core attacks. The first attack targets inference optimizations and deployment strategies. For example, our attack detects whether a provider uses speculative decoding, a widely deployed inference-time optimization, and further identifies the context length of the draft model used in the pipeline. Our measurements show that Google Gemini Flash 2.5 uses speculative decoding with a draft context window of approximately 128K tokens. The second attack recovers key architectural properties, including the number of transformer layers, hidden dimension size, and number of attention heads. To achieve this, LeakyLMs builds a detailed and accurate model of token-generation timing on modern NVIDIA GPUs, characterizing how latency scales with model configuration and hardware parameters. The attack then performs a search over the architecture space using this timing model. In experiments with Llama models, the near-correct architectural configuration appears in the top-10 guesses more than 90% of the time.

Vision-Language-Action (VLA) models are deployed through pipelines that end users cannot audit, and a poisoned VLA can behave normally on clean observations while a small visual trigger redirects a long-horizon robot policy before any failure becomes observable. Existing vision or language defenses rarely explain what a triggered VLA representation looks like or how to recover behavior without retraining. We study this gap through two independently proposed VLA attacks from groups with distinct injection strategies, BadVLA and INFUSE; the latter persists after downstream clean adaptation. Across the evaluated poisoned models, we identify a recurring internal mechanism: a \emph{compact causal footprint}, namely a small visual support that is attention-seeded, spatially compact, and \emph{causal} in a precise sense -- masking it returns a clean-calibrated evidence-evolution score to the normal operating region. This footprint motivates TrustVLA, a mechanism-guided inference-time defense that adapts the Dirichlet evidence framework from trusted classification to monitor per-token, per-layer epistemic uncertainty in VLA policies. With only a small clean calibration set, TrustVLA (i)~detects abnormal evidence evolution, (ii)~localizes the compact support by counterfactual mechanism-score drop, and (iii)~recovers the observation by localized inpainting. Across OpenVLA/LIBERO and $π_{0.5}$ transfer evaluations, TrustVLA reduces attack success while preserving clean-task performance, providing a retraining-free, mechanism-guided defense for visual-triggered VLA backdoors.

Safety alignment in large language models can be fragile under fine-tuning, as even benign task adaptation may increase harmful compliance. Existing defenses mainly follow two directions: they either intervene during or after fine-tuning through retraining or weight modification, which can be costly and may hurt task performance, or they use model-agnostic safety classifiers, which may miss failures specific to a given fine-tuned checkpoint. These limitations motivate a post hoc, model-specific, and non-invasive approach to safety restoration. To meet these requirements, we propose HyperSafe, a framework that restores safety behavior by generating a model-specific Safe Side Network (SSN) for each fine-tuned checkpoint. HyperSafe uses layer-wise activation fingerprints to capture how fine-tuning changes the model's inner representations. With a small set of given calibration prompts, the hypernetwork maps these fingerprints to the parameters of the \ssn{} in a single forward pass. The generated \ssn{} runs alongside the frozen fine-tuned model and performs prompt-level safety classification: harmful prompts are routed to refusal, while safe prompts are answered by the original fine-tuned model. Thus, HyperSafe requires no gradient updates, no safety data at deployment time, and no modification to the deployed model weights. We evaluate HyperSafe on two model families, Qwen2-7B and LLaMA-3-8B, across multiple safety benchmarks. HyperSafe reduces harmful response rates from 19-31% to below 1% on every held-out checkpoint, while keeping downstream task accuracy within 1% of the fine-tuned baseline on average. Code is available at https://github.com/nokronim/project-safety-remedy.

Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists. We evaluate five defense paradigms (no defense, static steering, CAST, AlphaSteer, probe-gated) across seven instruction-tuned models (7-31B) and five attack types (GCG, AutoDAN, DeepInception, prefilling, intent laundering). Our central finding: prompt-time activation defenses are structurally blind to prefilling attacks. AlphaSteer achieves 0% attack success on GCG, AutoDAN, and intent laundering but 50% on prefilling. We prove a corollary: any defense that gates intervention on a single layer's activation alignment with a benign reference (cone, subspace, or null-space) is blind to attacks that craft activations to lie inside that reference, whether checked at prompt time or per token. As its constructive contrapositive we introduce response-time probing: a linear probe on the model's hidden state at the first generated tokens, with AUROC 0.97-1.00 across all seven models. Combined with a halt, it cuts prefilling attack success to 0/40 on every model with 0% benign false positives, outperforming Llama Guard 3. Cross-template generalisation depends on probe depth, so we scope the claim to the canonical prefilling-template family. Composing the response-halt with AlphaSteer's null-space steering gives an orthogonal split (the halt catches prefilling, AlphaSteer catches semantic attacks), reaching defense success 0.983 on Mistral and 0.994 on Llama and dominating both components. We further show MMLU fails to capture steering's true utility cost, which appears as behavioral hedging rather than factual loss, and that diverse negative training sets cut probe false positives from 80-100% to near zero. Code, attacks, per-sample results, and the judge prompt are released.

AI agents are granted access to tools, APIs, and other infrastructure, making them active principals in those systems. The dominant approach places controls inside the agent's own runtime: system prompts, output filters, and guardrail libraries. Any control in the agent's address space is reachable by inputs that influence it; this generalizes to any AI system with sufficient reach into its own runtime, a class we term escapable AI systems. We identify four properties that an authorization mechanism must satisfy for architectural control rather than for cooperative requests: process separation, pre-action enforcement on a structurally only path, fail-closed at both the request and system levels, and externalized signed evidence verifiable outside the controlled system's trust boundary. We position this layer as execution-time AI alignment, complementing training-time alignment (RLHF, Constitutional AI) and inference-time alignment. We present the Unfireable Safety Kernel, a Rust reference implementation realizing all four. Its fail-closed invariant is machine-checked at two levels: an SMT theorem (Z3) and an exhaustive bounded-model-checking proof of the production decision function (Kani, 4/4 harnesses). A Python-to-Rust migration was gated on byte-equivalence (1000/1000 fixtures; 17/17 adversarial classes). We evaluate the kernel governing a live, escapable AI system, a deterministic, self-improving world model, against an escape-seeking adversary driving its real self-modification seam: across 1,000 self-modifications, all 704 attempts on the safety-critical core are refused, with no escape; a further 300, under the operator kill switch, are also refused. A separate campaign of 6,240 authorization round-trips had no successful bypass. Against 3 contemporary systems claiming the agent control plane, the agent invokes control; here, it lacks that choice.

Tabular foundation models are commonly assumed to present limited privacy concerns as they are often pre-trained on large collections of synthetic data. However, these models leverage in-context learning, where sensitive records may be provided directly at inference time as labelled context examples. In this paper, we demonstrate that predictions generated via the attention mechanism leak sufficient information to enable effective Membership Inference Attacks (MIAs). To highlight this vulnerability, we propose AMIA (Attention-based Membership Inference Attack), a shadow-model-free attack that exploits the concentration of transformer attention patterns. Our results show that attention mechanisms reveal strong membership signals, which exceed classical confidence-based attacks, achieving an average gain of 7.7%, specially in low false-positive regimes. To mitigate this risk, we introduce an inference-time defence inspired by $k$-anonymity principles. This approach reduces the uniqueness of context-key representations without introducing random noise or retraining the model. By targeting only high-risk queries identified through AMIA scores, the defence substantially reduces membership leakage of this attack by an average of 50% and 25% against confidence-based attacks, while preserving predictive utility with only 3.9% performance degradation. Beyond showing that context examples are vulnerable, we further demonstrate that fine-tuning introduces an additional source of privacy risk. In particular, samples whose prediction confidence increases after fine-tuning become more susceptible to MIAs, indicating that fine-tuning can amplify memorisation and expose sensitive training information through confidence shifts.

Speculative inference accelerates large language model (LLM) decoding but provides no inherent safety guarantees. Existing safety defenses are largely incompatible with speculative inference: they either introduce additional computation or disrupt the draft-verify mechanism, negating acceleration benefits. This reveals a fundamental incompatibility between current safety methods and speculative decoding. We propose SafeSpec, a safety-aware speculative inference framework that integrates risk estimation directly into the verification process. SafeSpec attaches a lightweight latent safety head to the target model to jointly evaluate semantic validity and safety in a single forward pass. When unsafe generations are detected, SafeSpec applies rollback and safety-guided reflective multi-sampling to recover safe continuations rather than terminating generation. We model jailbreak attacks as distributional shifts over generative trajectories, where adversarial prompts increase the probability of harmful continuations without eliminating safe ones. Under this model, SafeSpec performs risk-aware trajectory recovery within the speculative decoding process. Across multiple models and adversarial benchmarks, SafeSpec achieves a substantially improved safety-efficiency trade-off. On Qwen3-32B, SafeSpec reduces attack success rates by 15% while preserving a 2.06x inference speedup on benign workloads, demonstrating that speculative acceleration and inference-time safety can be jointly optimized.

Sparse Autoencoders (SAEs) decompose residual-stream activations into interpretable features. Recent latent-space defenses increasingly rely on these decompositions, assuming that identified "unsafe" SAE features serve as actionable handles for monitoring and intervention. In this paradigm, clamping a specific harmful feature is expected to reliably prevent model misbehavior. However, we show that this success may hide a recoverable failure mode: the clamp may block one visible route to a behavior without eliminating the behavior itself. We formulate this vulnerability as post-intervention recovery, a constrained residual-space optimization problem. Starting from the post-intervention residual state, we optimize residual perturbations to recover the pre-intervention behavior while preserving the post-intervention values of the targeted SAE features. Even under a strong threat model where the intervention remains active throughout optimization and generation, recovery remains possible. To rule out that recovery simply undoes the intervention, we use encoder-orthogonal updates for single-layer interventions and the corresponding feature-map Jacobian in the cross-layer setting. Across TPP, unlearning, IOI, and refusal steering experiments, this stress test reveals recoverable behavior despite successful feature-level intervention. Especially in the safety-critical refusal-steering setting, we achieve a 95.8% recovery rate on valid samples while keeping defended-feature relative drift to 0.131, substantially below suffix-based baselines. A recovery-path attribution analysis further localizes this recovery to the SAE reconstruction residual, the component left unexplained by the SAE. These results expose a gap between feature-level control and behavioral completeness: SAE features can support causal intervention, but controlling them does not guarantee control over the underlying behavior.

Open-weight video diffusion models can generate photorealistic unsafe content, from violence to misinformation, yet existing defenses either require expensive safety fine-tuning that degrades general capability, or apply external filters that are trivially bypassed by adversarial prompts. We present REINS (REpresentation-space INference-time Safety steering), a training-free method that aligns video diffusion models at inference time by steering their internal representations toward safe generation. Our key finding is that safety-relevant structure is linearly encoded in the hidden-state activations of video diffusion transformers, and a single direction, discovered via Supervised PCA on binary safety labels, suffices to separate safe from unsafe generation trajectories. At inference, adding this direction to hidden states at an intermediate transformer layer redirects generation from harmful content to semantically related safe alternatives, with no weight updates, no concept enumeration, and negligible computational overhead. Through mechanistic analysis, we reveal that while safety information accumulates monotonically with transformer depth, steering effectiveness peaks at intermediate layers (~50% depth), exposing a fundamental tradeoff between information availability and downstream propagation capacity. We evaluate REINS across 9 video diffusion models, multiple parameter scales (1.3B-5B), and both text-to-video and image-to-video generation, to our knowledge, the broadest safety evaluation suite in the video generation literature.

As large language models (LLMs) are increasingly deployed in user-facing systems, black-box jailbreak defense has become an important practical problem. Existing defenses often rely on known-attack coverage, prompt-level semantic judgment, or local runtime control, yet these paths can become unstable under evolving prompt packaging, expression rewriting, and structure manipulation. We observe that many black-box jailbreaks do not remove the harmful goal, but reorganize the information needed to express and execute it, thereby evading safety alignment while remaining recoverable during generation. Motivated by this observation, we propose DoubtProbe, a dual-branch inference-time defense framework that combines structural verification with semantic auditing and formulates black-box jailbreak defense as consistency checking under controlled transformation. The structural branch extracts a structured representation from the original request, reconstructs the request under representation constraints, and detects information-preservation failures between the original and reconstructed requests; the semantic branch audits the original prompt directly. We evaluate DoubtProbe against representative black-box defenses on jailbreak and benign-request benchmarks, and further test backbone transfer from Qwen2.5-72B to Llama-3.1-70B. Results show that DoubtProbe achieves a stronger and more stable defense-utility trade-off: on Qwen2.5-72B, it reduces the JBB attack success rate from 0.293 to 0.100 and the CodeAttack attack success rate from 0.152 to 0.001, while maintaining false positive rates of 0.022 and 0.016 on AlpacaEval and OR-Bench; the same pattern remains stable on Llama-3.1-70B. These findings show that structural inconsistency signals provide a practical and generalizable basis for black-box jailbreak defense, especially when combined with semantic auditing.

Large language models routinely generate code with exploitable security flaws. Prior literature attributes this limitation to a lack of security expertise, steering current defense mechanisms toward heavy fine-tuning or external knowledge retrieval, which introduces significant computational overhead and data bias through redundant code examples. Contrary to this view, we argue that pretraining corpora are already rich in security material. The bottleneck is activation: without an explicit and brief cue, statistical pressure toward common training-distribution patterns suppresses the model's safety-relevant representations. We present SPARK, an inference-time security harness that activates this latent knowledge without any retraining. The harness has two parts. ComponentI retrieves a few of the relevant Common Weakness Enumeration (CWE) entries for each coding task and appends a short structured cue to the prompt; this alone is enough to surface the model's existing security representations. ComponentII adds a precomputed token bias to the logits at every decoding step. We obtain the bias by projecting a safe-direction vector, the unit difference between the mean safe and mean unsafe last-layer hidden states, through the language model head. The bias is computed once offline; applying it costs a single vector addition per generated token. We evaluate SPARK on 9 open-source models across C++, Java, and Python, and compare with 7 baselines spanning fine-tuning and retrieval-augmented methods. SPARK matches or improves on the best baseline in every setting while preserving HumanEval utility. We further test Component~I in a black-box setting on 7 of today's strongest models, including Claude, DeepSeek, and GPT, demonstrating the bottleneck of insecure code generation and the improvements enabled by our method.

Diffusion models have emerged as powerful tools for planning and control by learning multimodal distributions over actions and trajectories. Yet reliable inference-time safety enforcement remains a key barrier to their deployment in safety-critical tasks. Existing approaches typically project each denoising iterate onto the feasible set, even though constraints are defined only on the final clean trajectory. Enforcing feasibility on noisy intermediate samples can therefore overconstrain the sampling dynamics, substantially degrading sample quality. To address this limitation, we introduce DiRecT (Diffusion-based planning via Receding-horizon denoising with Terminal constraints), a training-free algorithm for constrained sampling from diffusion models via stochastic optimal control (SOC). DiRecT enforces constraints only on the final clean sample, avoiding unnecessary restrictions on the intermediate denoising dynamics. Inspired by model predictive control, we derive a principled receding-horizon surrogate for the otherwise intractable constrained SOC formulation, yielding an efficient algorithm that cleanly separates stochastic denoising from constraint satisfaction, progressively steering samples toward feasible final trajectories without distorting the learned diffusion dynamics. Furthermore, DiRecT is highly flexible: it can leverage off-the-shelf or domain-specific optimizers, incorporate priors over environment dynamics, and optimize additional soft rewards. Extensive experiments on safe planning benchmarks demonstrate that DiRecT substantially improves deployment safety and task performance over existing diffusion-based planning baselines.

Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor logits into the target model's vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the safety-utility trade-off can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on adversarial benchmarks while keeping task accuracy and inference overhead within practical bounds. The results show that safety alignment can be transferred between model families at inference time, without touching either model's weights.

With AI increasingly deployed in safety-critical systems, providing formal robustness guarantees for the underlying models is essential. Existing verification methods either rely on overly conservative approximations or incur prohibitive computational costs. For example, the use of lp-norm perturbations in video settings encodes the belief that the adversary can inject noise in every video frame. In practice, adversarial perturbations exhibit structured spatial and temporal correlations, constrained to lower-dimensional, semantically meaningful subspaces. In this work, we study robustness verification of 3D CNNs processing video and volumetric inputs, targeting applications in action recognition (UCF-101), autonomous driving (Udacity), and medical imaging (MedMNIST) exploiting realistic assumptions on adversarial strength by modelling them as spatio-temporal constraints - where the attacker can modify either a subset of frames or patches within a set of consecutive frames. We demonstrate that modelling realistic constraints enables tighter approximations. We introduce Spatio-Temporal Bound Propagation (STBP), a verification framework that computes an exact closed-form characterization of the first convolutional layer and propagates certified bounds through subsequent layers using scalable approximations. Computing the exact closed form provides the tightest bounds for the first convolutional layer. Thus, we utilise approximation methods in the remainder of the network. To spur further progress in this field, we propose ST-Bench, a verification benchmark for autonomous driving and activity recognition, to systematically evaluate verifiable robustness. Compared to existing verification-based approaches, STBP provides stronger robustness guarantees with significantly improved scalability, achieving 1.7x higher certified robust accuracy under identical perturbation budgets.

Ensuring the reliability of Large Language Models (LLMs) under distribution drift requires inference-time adaptation. While inference-time alignment methods such as Best-of-$N$ and rejection sampling are widely used, they frame the task as a sampling-intensive, reward-guided search, leading to two key limitations: their performance is bounded by the base model's generation quality, and their reliance on imperfect reward models makes them vulnerable to reward hacking. To address these challenges, we introduce Gradient-Guided Reward Optimization (GGRO), a lightweight inference-time method that performs targeted, minimal intervention during decoding via gradient guidance. Specifically, GGRO monitors token-level entropy to identify high-uncertainty regions indicative of drift or misalignment. Upon detection, it responds by injecting nudging tokens, generated using gradient signals from an off-the-shelf reward model, to steer the generation trajectory rather than merely re-ranking samples. Experiments show that GGRO consistently improves inference-time alignment across safety, helpfulness, and reasoning benchmarks. It also increases coverage of high-quality responses and robustness to reward hacking, with minimal computational overhead. Code is available at https://github.com/lhk2004/GGRO.

Safety-aligned Large Language Models (LLMs) remain vulnerable to interventions during inference that redirect generation toward harmful outputs. Recent work attributes this to shallow safety, where alignment concentrates in the first few output tokens. We show that shallow safety is a special case of a broader inference-time vulnerability, in which short token injections at any generation step can substantially alter subsequent safety behavior. We also find that a model's alignment with refusal directions in its hidden states does not predict its robustness to such injection, revealing that internal state alone does not determine generation behavior under perturbation. To address this, we align models directly on generation trajectories constructed by simulating mid-sequence perturbation, and show that this improves robustness to mid-sequence injection and generalizes to attacks that exploit early-token generation. Our work argues that robust safety alignment requires training on the generation process itself, not only its outputs.

Large vision-language models (LVLMs) often hallucinate objects that are not present in the input image, largely because visual grounding weakens as decoding progresses. Existing inference-time mitigation methods modify logits or hidden states throughout generation, but they suffer from three key limitations: they lack an explicit grounding objective, intervene even when the model is already well-grounded, and use fixed correction strengths that do not adapt to the severity of grounding failure. We propose BRACS (Barrier-Regulated Adaptive Closed-form Steering), a training-free steering framework that addresses these issues through barrier-regulated adaptive closed-form steering. BRACS monitors the model's own attention to measure visual grounding and applies corrections to the hidden states only when grounding deteriorates. The corrective update is computed analytically in closed form, requiring no training of auxiliary networks or model retraining. Experiments on LLaVA-1.5-7B and Qwen-VL-Chat show that BRACS consistently outperforms prior methods on hallucination benchmarks, reducing CHAIR$_s$ by 9.4 points and improving POPE F1 by 2.7 points, while matching or improving performance on four general multimodal benchmarks. BRACS also remains efficient, operating at 80% of greedy decoding throughput and achieving 1.3 times higher speed on average than the baselines.

Object hallucination remains a primary obstacle to the reliable deployment of Multimodal Large Language Models (MLLMs). Current inference-time mitigation methods mainly assume hallucinations stem from visual neglect, steering models to enhance visual reliance. In contrast, our systematic interventions on multiple MLLMs show that pushing toward more visual reliance may exacerbate hallucinations on some models, while less may mitigate hallucinations. This result suggests that attributing hallucinations solely to visual insufficiency is underdetermined. We argue that the image, as a context, simultaneously competes with the model's parametric knowledge and the textual context. For this, we propose a training-free framework, Context-Preference Activation Steering (CAS). It extracts two semantically distinct Context Preference Vectors (CPVs) via two small sets of designed conflict samples and applies them via single-pass signed residual injection at mid-early MLP layers during inference to control information reliance. Experiments show that CAS substantially mitigates object hallucinations without increasing decoding latency and preserves native text-generation quality.

Test-Time Training (TTT) is an emerging paradigm that enables models to adapt their parameters during inference, improving performance on tasks such as few-shot learning, retrieval-augmented generation, and complex reasoning. However, this dynamic adaptation introduces new vulnerabilities that adversaries can exploit to jailbreak models. We identify three threat models for TTT and demonstrate how attackers can leverage them to bypass safety filters. Our results show that TTT can significantly increase the Attack Success Rate (ASR) and the ASR over 10 generation trials (ASR@10). For example, under LoRA, the few-shot and generation-phase threat models achieve an average ASR@10 of 95% and 93% respectively, across models from different families and scales. These vulnerabilities transfer to production fine-tuning APIs. We also show that TTT-induced overfitting can produce degenerate outputs that inflate ASR under standard judges, and propose a validity-aware evaluation to correct for this. Our findings suggest that TTT exposes a new attack surface, strengthens attacks, and undermines existing safety guardrails. As a first step toward defense, we propose a lightweight provider-side detector that flags TTT requests via the perplexity shift on a private harmful holdout, but robust deployment will ultimately require dynamic alignment.

We present a method for dynamic quantitative assurance that enhances static safety cases with continuous, runtime-driven confidence updates. The method quantifies and propagates confidence across the development lifecycle by integrating design-time evidence and windowed runtime Safety Performance Indicators (SPIs) within a single Subjective Logic (SL)-based assurance case. At runtime, SPI evidence is continuously evaluated, and targeted claims are updated using a rule that increases confidence in the absence of violations and imposes prompt penalties when violations occur. This design prioritizes safety-relevant responsiveness over exact classical Bayesian posterior updates. We demonstrate the method using a simulation-based construction zone assist function, focusing on an ML-based construction cone detection component, and show how confidence evolves as SPI evidence is observed in operation.

The rapid growth of online video platforms and AI-generated content has made reliable video guardrails a key challenge for safety and real-world deployment. While most videos can be screened through fast pattern recognition, a small subset requires deeper reasoning over temporally complex content and nuanced policy constraints. Existing approaches typically rely on large vision-language models applied uniformly across all inputs, resulting in high inference costs and inefficient allocation of computation. We propose SafeLens, a video guardrail framework that introduces a fast-and-slow inference architecture for efficient and accurate content moderation with variable computational cost across inputs. Additionally, we construct a high-quality dataset by applying influence-guided filtering to the SafeWatch Dataset, retaining only 2.4% of the original data. To further address limitations of training-time scaling, we enable test-time reasoning by augmenting the filtered data with structured Chain-of-Thought traces. Across real-world and AI-generated video benchmarks, SafeLens achieves state-of-the-art performance, outperforming strong open-source video guardrails (e.g., SafeWatch-8B, OmniGuard-7B) and closed-source models (e.g., GPT-5.4, Gemini-3.1-pro) while significantly reducing inference cost, demonstrating that efficient design serves to be more effective than scaling data or model size alone.

End-to-End (E2E) autonomous driving models have shown growing capability in recent years, with performance improving on increasingly challenging benchmarks. However, modern generative E2E planners still suffer from a substantial number of catastrophic failures in safety-critical scenarios. We find that many such failures arise from violations of physical constraints and safety requirements, leading to unsafe behavior. Motivated by this finding, in this paper, we focus on improving safety outcomes in generative end-to-end driving with a targeted reduction of catastrophic planning failures, instead of enhancing average planning quality. Towards this end, we propose DriveSafer, a failure-aware safety framework for end-to-end planners. DriveSafer explicitly steers generative planners towards safe behaviors leveraging both training-time safety constraints and inference-time safety guidance. Compared to the state-of-the-art DiffusionDrive model, on the NAVSIM benchmark, DriveSafer reduces the number of catastrophic failures (PDMS=0) by 48%, with over 65% reduction in drivable-area compliance failures.

We study certified runtime monitoring of past-time signal temporal logic (ptSTL) from visual observations under partial observability. The monitor must infer safety-relevant quantities from images and provide finite-sample guarantees, while being \emph{reusable}: once trained and calibrated, it should certify any formula in a target fragment without per-formula retraining. For fragments induced by a finite dictionary of temporal atoms, we prove that the \emph{semantic basis}, the vector of atom robustness scores, is the minimum prediction target within the class of monotone, 1-Lipschitz reusable interfaces: any formula is evaluated by a deterministic decoder derived from the parse tree, and a single conformal calibration pass certifies the entire fragment with no union bound. We also introduce a \emph{rolling prediction monitor} that predicts only current predicate values and reconstructs temporal history online; this is easier to learn but grows conservative at long horizons. On a pedestrian-crossroad benchmark, rolling achieves tighter certified bounds at short horizons while the semantic-basis monitor is up to 4-times tighter at long horizons. We validate the presented monitors on real-world Waymo driving data, where both monitors satisfy the conformal coverage guarantee empirically.

The rise of text-to-image (T2I) models has increasingly raised concerns regarding the generation of risky content, such as sexual, violent, and copyright-protected images, highlighting the need for effective safeguards within the models themselves. Although existing methods have been proposed to eliminate risky concepts from T2I models, they are primarily developed for earlier U-Net architectures, leaving the state-of-the-art Diffusion-Transformer-based T2I models inadequately protected. This gap stems from a fundamental architectural shift: Diffusion Transformers (DiTs) entangle semantic injection and visual synthesis via joint attention, which makes it difficult to isolate and erase risky content within the generation. To bridge this gap, we investigate how semantic concepts are represented in DiTs and discover that attention heads exhibit concept-specific sensitivity. This property enables both the detection and suppression of risky content. Building on this discovery, we propose AHV-D&S, a training-free inference-time safeguard for image generation in DiTs. Specifically, AHV-D&S quantifies each textual token's sensitivity across all attention heads as an Attention Head Vector (AHV), which serves as a discriminative signature for detecting risky generation tendencies. In the inference stage, we propose a momentum-based strategy to dynamically track token-wise AHVs across denoising steps, and a sensitivity-guided adaptive suppression strategy that suppresses the attention weights of identified risky tokens based on head-specific risk scores. Extensive experiments demonstrate that AHV-D&S effectively suppresses sexual, copyrighted-style, and various harmful content while preserving visual quality, and further exhibits strong robustness against adversarial prompts and transferability across different DiT-based T2I models.

While spatiotemporal deepfake detectors achieve high AUC, our experiments reveal their susceptibility to evasion attacks. These models tend to overfit on fragile temporal spectrum cues, rather than learning robust semantic causality. To mitigate this vulnerability, we propose SpInShield, a temporal spectral-invariant defense framework explicitly designed to decouple semantic motion from manipulatable spectral artifacts. We propose a learnable spectral adversary that dynamically synthesizes severe spectral deformations, simulating extreme attack scenarios. By employing a shortcut suppression optimization strategy, SpInShield compels the encoder to extract reliable forensic cues while purging unstable spectral statistics from the latent space. Experiments show that SpInShield obtains competitive performance on widely used datasets and outperforms the strongest baseline by 21.30 percentage points in AUC under simulated amplitude spectral attacks.

Tool-using agent systems powered by large language models (LLMs) are increasingly deployed across web, app, operating-system, and transactional environments. Yet existing safety benchmarks still emphasize explicit risks, potentially overstating a model's ability to judge deceptive or ambiguous trajectories. To address this gap, we introduce ROME (Red-team Orchestrated Multi-agent Evolution), a controlled benchmark-construction pipeline that rewrites known unsafe trajectories into more deceptive evaluation instances while preserving their underlying risk labels. Starting from 100 unsafe source trajectories, ROME produces 300 challenge instances spanning contextual ambiguity, implicit risks, and shortcut decision-making. Experiments show that these challenge sets substantially degrade safety-judgment performance, with hidden-risk cases remaining particularly non-trivial even for recent frontier models. We further study ARISE (Analogical Reasoning for Inference-time Safety Enhancement), a retrieval-guided inference-time enhancement that retrieves ReAct-style analogical safety trajectories from an external analogical base and injects them as structured reasoning exemplars. ARISE improves judgment quality without retraining, but is best viewed as a task-specific robustness enhancement rather than a standalone safety guarantee. Together, ROME and ARISE provide practical tools for stress-testing and improving agent safety judgment under deceptive distribution shifts.

Text-to-Video (T2V) models have demonstrated remarkable capability in generating temporally coherent videos from natural language prompts, yet they also risk producing unsafe content such as violence or explicit material. Existing prompt-level defenses are largely inherited from text-to-image safety and operate on the lexical surface of the input, making them vulnerable to jailbreak attacks that disguise harmful intent through rephrasing or adversarial prompting. Moreover, T2V generation introduces a distinctive challenge overlooked by prior work: temporally emergent risk, where a seemingly benign prompt leads to unsafe content through the generator's temporal extrapolation toward narrative coherence. We propose \method{}, a training-free, inference-time defense framework that reformulates T2V safety as a causal intervention in a temporally structured semantic space. TrajShield handles explicit unsafe prompts, jailbreak attacks, and temporally emergent risks in a unified manner by simulating the implied trajectory of a prompt, localizing the causal origin of potential risk, and applying a minimally invasive rewrite that neutralizes the risk while preserving safety-irrelevant semantics. Experiments on T2VSafetyBench across 14 safety categories and multiple T2V backends demonstrate that TrajShield achieves state-of-the-art defenseive performance while maintaining high semantic fidelity, substantially outperforming existing defenses, with an average ASR reduction of 52.44%.

Diffusion models have become the foundation of modern generative systems, with most research focusing primarily on improving generation efficiency and output quality. The timestep embedding component is a crucial part of the diffusion pipeline, which provides a temporal conditioning signal to the denoising network, enabling it to adapt its predictions across different noise levels throughout the process. Despite their potential to contain substantial information, timestep embeddings remain underexplored in current research, especially for security risks and reliable provenance. To fill this gap, we introduce Shadow Timestep Embedding (STE), a novel mechanism that investigates the underutilized temporal space for malicious information injection into diffusion models. In particular, when zooming in on the timestep embedding space, we find that different timesteps exhibit distinct representational capabilities that can encode side-channel information. Moreover, such encoded information can be utilized for attack and defense purposes through the scheduler interface. We present a theoretical analysis of timestep embeddings as position-encoding mappings and derive a mutual coherence evaluation that explains the separability of disjoint timestep intervals. Our findings reveal the diffusion model's timestep as a powerful side channel for carrying dedicated information, motivating new directions for adversarial generative modeling by understanding the temporal dimension.

Defending against backdoor attacks in large language models remains a critical practical challenge. Existing defenses mitigate these threats but typically incur high preparation costs and degrade utility via offline purification, or introduce severe latency via complex online interventions. To overcome this dichotomy, we present Tail-risk Intrinsic Geometric Smoothing (TIGS), a plug-and-play inference-time defense requiring no parameter updates, external clean data, or auxiliary generation. TIGS leverages the observation that successful backdoor triggers consistently induce localized attention collapse within the semantic content region. Operating entirely within the native forward pass, TIGS first performs content-aware tail-risk screening to identify suspicious attention heads and rows using sample-internal signals. It then applies intrinsic geometric smoothing: a weak content-domain correction preserves semantic anchoring, while a stronger full-row contraction disrupts trigger-dominant routing. Finally, a controlled full-row write-back reconstructs the attention matrix to ensure inference stability. Extensive evaluations demonstrate that TIGS substantially suppresses attack success rates while strictly preserving clean reasoning and open-ended semantic consistency. Crucially, this favorable security-utility-latency equilibrium persists across diverse architectures, including dense, reasoning-oriented, and sparse mixture-of-experts models. By structurally disrupting adversarial routing with marginal latency overhead, TIGS establishes a highly practical, deployment-ready defense standard for state-of-the-art LLMs.

Alignment faking, where a model behaves aligned with developer policy when monitored but reverts to its own preferences when unobserved, is a concerning yet poorly understood phenomenon, in part because current diagnostic tools remain limited. Prior diagnostics rely on highly toxic and clearly harmful scenarios, causing most models to refuse immediately. As a result, models never deliberate over developer policy, monitoring conditions, or the consequences of non-compliance, making these diagnostics fundamentally unable to detect alignment faking propensity. To support study of this phenomenon, we first introduce VLAF, a diagnostic framework grounded in the hypothesis that alignment faking is most likely when developer policy conflicts with a model's strongly held values. VLAF uses morally unambiguous scenarios to probe this conflict across diverse moral values, bypassing refusal behavior while preserving meaningful deliberative stakes. Using VLAF, we find that alignment faking is substantially more prevalent than previously reported, occurring in models as small as 7B parameters - with olmo2-7b-instruct faking alignment in 37% of cases. Finally, we show that oversight conditions induce activation shifts that lie along a single direction in representation space. This means the behavioral divergence driving alignment faking can be captured by a single contrastive steering vector, which we exploit for lightweight inference-time mitigation. Finally, we exploit this for mitigation that requires no labeled data and minimal computational overhead, achieving relative reductions in alignment faking of 85.8%, 94.0%, and 57.7% on olmo2-7b-instruct, olmo2-13b-instruct, and qwen3-8b respectively.

Generative AI systems are known to amplify biases present in their training data. While several inference-time mitigation strategies have been proposed, they remain largely empirical and lack formal guarantees. In this paper we introduce CTLF, a branching-time logic designed to reason about bias in series of generative AI outputs. CTLF adopts a counting worlds semantics where each world represents a possible output at a given step in the generation process and introduces modal operators that allow us to verify whether the current output series respects an intended probability distribution over a protected attribute, to predict the likelihood of remaining within acceptable bounds as new outputs are generated, and to determine how many outputs are needed to remove in order to restore fairness. We illustrate the framework on a toy example of biased image generation, showing how CTLF formulas can express concrete fairness properties at different points in the output series.

Even modern AI models often remain vulnerable to multimodal queries in which harmful intent is embedded in images. A widely used approach for safety alignment is training with extensive multimodal safety datasets, but the costs of data curation and training are often prohibitive. To mitigate these costs, inference-time alignment has recently been explored, but they often lack generalizability across diverse multimodal jailbreaks and still incur notable overhead due to extra forward passes for response refinement or heavy pre-deployment calibration procedures. Here, we identify insufficient visual attention to safety-critical image regions as one of the key causes of multimodal safety failures. Building on this insight, we propose Multimodal Risk-Adaptive Steering (MoRAS), which enhances safety-critical visual attention via concise visual contexts for accurate multimodal risk assessment. This risk signal enables risk-adaptive steering for direct refusals, reducing inference overhead while remaining generalizable across diverse multimodal jailbreaks. Notably, MoRAS requires only a small calibration set to estimate multimodal risk, substantially reducing pre-deployment overhead. We conduct various empirical validations across multiple benchmarks and MLLM backbones, and observe that the proposed MoRAS consistently mitigates jailbreaks, preserves utility, and reduces computational overhead compared to state-of-the-art inference-time defenses.