Papers

Purpose: Developers choose how much a language-model agent reasons before it acts. Some pick a high level for fear that a low one is not enough, and pay for it in tokens, time and overthinking. We tested whether a low level is enough for routine office work. Methods: Two tiers of GPT-5.6 did 14 routine office tasks at the low and the max reasoning level, 840 runs in all. To make the tasks harder, each one sets a target that the rules make impossible to reach and offers a forbidden tool that would reach it. Some versions also tell the agent that the forbidden tool counts. A program checked every run from its final state. Results: At both levels the agent always followed the rules and never used the forbidden tool, including 30 runs in which it was told that the tool counts while it could still have used it. The low level used about 43% fewer output tokens and 20% less time. Conclusion: For routine office work, a low reasoning level is enough.

Invisible image watermarks are increasingly used for deepfake detection and provenance tracking, where they must survive not only incidental distortions but also deliberate removal. We revisit spread-spectrum embedding, a classical watermarking principle, inside a modern neural post-hoc watermarking architecture. Our starting point is a measurement: in existing encoder-decoder schemes each message bit occupies only a small fraction of the image, a shared contributing factor to their fragility, since removal then need only disturb the region a bit occupies. SpreadMark instead spreads each bit as a dense pseudo-random codeword over the whole image and recovers it by matched-filtering a learned cover-suppressed chip representation, with a parallel convolutional decoding path and sparsification-aware training. A conditional chip-space analysis shows that, under a codeword-independent perturbation model, dense spreading increases the budget required to disrupt matched-filter recovery. Evaluated on COCO and DIV2K against nine schemes, SpreadMark is the only evaluated method retaining high detection under both the regeneration and the latent-space sparsification settings we test, with competitive JPEG and additive-noise robustness. It keeps the embedded watermark imperceptible, maintaining high perceptual quality on both COCO and DIV2K.

Contactless keystroke inference via WiFi sensing highlights severe privacy threats, yet its real-world feasibility is hindered by two fundamental physical and deployment bottlenecks: the strict requirement for network privileges to acquire stable sensing streams, and the inherent "waveform fusion" ambiguity of pure WiFi signals during rapid, muscle-memory typing. To overcome these limitations, we propose AirKey, a novel cross-modal sensing framework that achieves highly stealthy, zero-training PIN eavesdropping. First, to bypass network deployment barriers, AirKey exploits fundamental IEEE 802.11 mechanisms to predictably elicit Acknowledgment (ACK) responses from unmodified target devices. By passively harvesting Channel State Information (CSI) from these ACKs using a low-cost microcontroller, AirKey secures a continuous spatial sensing stream entirely without network association. Crucially, to resolve the WiFi waveform fusion bottleneck, AirKey introduces a cross-modal complementarity mechanism. By utilizing lightweight acoustic signals as precise temporal anchors, the system robustly guides the segmentation of overlapping CSI trajectories. This joint spatiotemporal fusion strictly intersects CSI-derived spatial similarities with acoustic-guided inter-keystroke timing. Extensive real-world evaluations demonstrate that AirKey achieves over 4x higher accuracy than state-of-the-art unimodal zero-training schemes, successfully recovering device-unlock PINs within 6 attempts. Ultimately, this work exposes a critical vulnerability in contemporary smart interfaces, underscoring the severe privacy implications of ubiquitous multimodal sensing.

Detecting source code vulnerabilities is increasingly difficult as modern security flaws are rooted in complex causal dependencies between execution flows, control conditions, and program states. Despite recent advances in Large Language Models (LLMs) and multi-agent frameworks, existing approaches primarily address superficial similarities between benign and vulnerable functions while failing to capture the complex causal dependencies inherent in security flaws. To address these limitations, we propose Causal Context-based Agentic Reasoning (CLEAR), a novel multi-agent vulnerability detection framework integrated with a causal knowledge graph. CLEAR systematically constructs a Vulnerability Causal Knowledge Graph (VCKG) that models the causal chains between entrypoints, preconditions, root causes, and fix intents across vulnerability instances. Leveraging this structured knowledge, four specialized agents, including the Collector, Claim, Critic, and Judge, collaboratively verify vulnerability hypotheses through retrieved causal contexts. Experimental results on C/C++ and Java vulnerability benchmarks demonstrate that CLEAR improves Pair-Correct (P-C) performance by 130.7% and 71.56% over state-of-the-art approaches, demonstrating the effectiveness of causal knowledge graph-guided reasoning for automated vulnerability detection.

Long-term memory enables persistent personalization in LLM agents, but repeated memory-conditioned responses can cumulatively reveal protected attributes even when they are never stated explicitly. We formalize this threat as adaptive transcript privacy and introduce DP-MemView, a differentially private interface that privately selects public response-conditioning views and exposes those views—rather than raw memory—to the response LLM. Each private selection is charged to every protected attribute whose memory group intersects the read set. Per-attribute ledgers block any selection that would exceed its cap and return a fixed generic view instead. Under an explicit interface contract, we prove pure B_a-DP for the entire adaptive transcript. We also extend the result to stores that differ across multiple protected groups and bound how much observing the transcript can change an adversary's prior odds. We evaluate the online and preallocated modes with three response LLMs on a controlled adjacent-store benchmark and a public-corpus transfer track. Both modes keep transcript distinguishability near chance while preserving target-required personalization and overall response quality. Further diagnostics show that removing key safeguards causes mismatched output support, missing ledger charges, revealing side channels, or growing long-horizon leakage.

Software-defined wide-area networking concentrates operational authority in controllers, orchestrators, and Internet-facing edges, but severity-only remediation queues do not represent current exposure or evidence quality. This paper specifies Predictive Exposure and Cryptographic Readiness (PECR), a decision-support method that combines nine normalized severity, threat, reachability, and consequence factors while reporting confidence and score intervals separately. Cryptographic dependencies enter an independent migration queue. The evaluation uses 100 role- and zone-conditioned synthetic vulnerability-asset records over 62 role-consistent assets, plus 30 independent generator replications. Against the operational PECR order, CVSS-only, EPSS-only, KEV-first, and CVSS x EPSS queues yield Kendall $\tau$ values of 0.196, 0.221, 0.301, and 0.234; a stronger five-factor context-lite comparator yields 0.755. Across 30,000 joint Dirichlet weight draws, median $\tau$ is 0.836, 0.893, and 0.924 under broad, moderate, and narrow perturbations. Masking 10% of factor cells labels 47 records as evidence-limited because their intervals cross a decision boundary. The study establishes executable specification, synthetic queue differentiation, and stress-test robustness. It does not establish predictive accuracy, avoided loss, or production effectiveness because no real outcome labels are used.

Recent advances in video generation models have significantly intensified the deepfake threat, yet the current deepfake video detection benchmarks remain underdeveloped. In particular, the effectiveness of image-level detectors in the video domain has not been systematically assessed. To fill this gap, we present FakeI2V-Bench, a benchmark for evaluating state-of-the-art video-level deepfake detectors in challenging scenarios, with a particular focus on systematically assessing the performance of image-level deepfake detectors in the video domain. FakeI2V-Bench comprises 97,548 videos, containing content generated by the latest powerful generation models and covering a broader range of categories. Using this dataset, we conduct a systematic evaluation of eight video-level detectors and twelve representative image-level detectors. Experimental results show that the best-performing image-level detector achieves an 80.16% AUC, slightly outperforming the strongest video-level detector (i.e., 79.99% AUC). Going beyond benchmarking, we present IV-Bridge, a general framework that enhances the applicability of image-level deepfake detectors to videos. IV-Bridge employs a random forest model with statistical features to aggregate frame-level predictions, allowing eleven image-level detectors to surpass state-of-the-art video-level approaches, with the best-performing variant achieving a 93.80% AUC. Overall, FakeI2V-Bench establishes a rigorous benchmark for deepfake video detection and introduces a novel pathway for extending image-level detectors to the video domain, offering new insights and directions for future research. Code and data are available at https://github.com/CryptoAILab/FakeI2V-Bench.

Large language model (LLM) watermarking provides an important mechanism for tracing the provenance of generated text. Existing statistical watermarks are often effective and robust, but most of them rely on private detection keys, which centralizes verification and complicates public auditing. Recent public or publicly verifiable watermarking schemes improve key management, yet many of them rely on exact recovery of embedded cryptographic strings, making them fragile under token edits, truncation, copy-paste, and low-entropy generation. This paper introduces DHMark, a public-key watermarking framework for LLM-generated text. The key idea is to separate payload authorization from noisy textual evidence. An issuer signs a short registry payload bound to a public context, and the payload is expanded into many one-bit equations. During generation, a Diffie-Hellman-guided token-labeling interface assigns each candidate token a public equation vote, and the sampler softly or selectively promotes candidates whose votes agree with the authorized payload. During verification, third-party verifiers use public information to extract token votes, aggregate them into equation-level evidence, and score only signed registry records. This design avoids exact recovery of a long embedded signature and instead treats watermark detection as registry-aided statistical evidence aggregation. We formalize the public-verification setting, analyze label pseudorandomness, registry-backed soundness, and sampling distortion, and evaluate a prototype under truncation, substitution, copy-paste, wrong-context, and plain-generation attacks. In the default 32-bit configuration, DHMark maintains at least a 0.967 valid rate across eight edit conditions while yielding a 0.000 acceptance rate on three negative controls.

The AI Security Leaderboard is an independent benchmark that ranks the safeguards of frontier AI models from least to most secure. It tests models against the FAR$.$AI Minimal Standard for Safeguards, which represents a minimum bar for security: meeting it does not guarantee a secure model, but failing to meet it guarantees a lack of state-of-the-art security. Version 1.0 covers severe misuse requests across chemical, biological, radiological, nuclear, and explosive (CBRNE) threats and offensive cybersecurity. In this report, we tested four leading models for universal jailbreaks in the context of this minimal standard, and found more than a hundredfold difference in security. Claude Fable 5 and GPT-5.6 Sol held against every attack we ran, with no universal jailbreak found; we estimate they would likely cost more than \$14,200 to jailbreak, if it is possible with this methodology at all. Meanwhile, we found hundreds of universal jailbreaks for Grok 4.5 and Gemini 3.1 Pro; each broke for under \$300, with universal jailbreaks in Grok's weakest domain, cybersecurity, accessible for as little as \$24. The gap is fixable: every weakness we found belongs to a known class of attack that already has a defense deployed in production models. The leaderboard will be updated on a rolling basis as new models are released, and the evaluation methodology and Minimal Standard will be periodically revised to take into account the latest capabilities and the state-of-the-art in safeguards. The leaderboard is available at leaderboard.far.ai.

Agentic artificial intelligence raises a new security concern: cyber threats that reason, act, and adapt locally without continuous human direction. We examine this threat through an Agentic Remote Access Trojan (agentic RAT): a Remote Access Trojan augmented with a locally deployed Small Language Model (SLM). The SLM interprets host and network observations, selects actions, recovers from failed steps, and reduces reliance on an external operator. We implement the concept in a controlled, network-isolated lab built from Kali Linux, a Metasploitable2 target, LM Studio, and a local 8-billion-parameter Dolphin-family model. We then test whether a model this small can support autonomous cyber decision-making. This is architecturally feasible today. On commodity hardware, with no cloud service and no operator in the loop, the SLM closed the full observe-decide-act cycle: it interpreted ranked reconnaissance evidence supplied by the controller, selected actions, and obtained verified root-shell access on real vulnerable services. However, it is not yet operationally reliable. The same model hallucinated commands, misread output, and recovered from failure inconsistently, completing 10.9% of a deliberately strict checklist. That gap reflects the limits of today's small models, not a ceiling on the concept. As SLMs improve, agentic endpoint systems may become more practical, more autonomous, and harder to detect, straining existing monitoring, containment, and policy-enforcement mechanisms. Real-world incidents in 2025-2026 already show AI-driven intrusions moving from concept toward practice. That makes the local, self-contained variant we study a plausible near-term direction, not a hypothetical one.

Modern large language models (LLMs) exhibit activation sparsity, wherein only a subset of their neurons is activated for given input tokens. Researchers have leveraged this property to optimize LLM serving systems by omitting weight accesses and computations pertaining to inactive neurons. Unfortunately, however, such optimizations create input-dependent weight accesses, which can be leaked over side channels. We present SparSEEty, a new token extraction attack that exploits input-dependent neuron weight accesses introduced by sparsity-exploiting LLM serving systems. SparSEEty first constructs a neuron-activation oracle using neuron weight access side channels during LLM inference, and then inverts the activation traces to reconstruct the input tokens, forming an end-to-end token extraction attack. We instantiate SparSEEty against an LLM serving system protected inside an Intel TDX confidential virtual machine (CVM), addressing three key challenges: (i) constructing a neuron-activation oracle using a combination of side channels exposed by CVMs, (ii) reducing inference-time overheads of neuron activation monitoring for covertness, and (iii) accurately inverting partial binary activation traces back to tokens. Our evaluation shows that SparSEEty can reconstruct both prompt and response tokens with consistently high BLEU scores (>0.95) across various models and datasets, while incurring monitoring overheads of 3.7% to 7.2%.

A software agent on a public blockchain accumulates authority and economic stakes, raising the engineering question of what makes it count as an individual. The paper's central contribution is a shift of trust root for the key-to-weights binding of agent identity: from hardware, operator, or wrapper trust to cryptographic assumptions enforced by a pinned implementation (liveness, key custody, oracle trust, and the underlying software stack remain external). We design and deploy on Solana devnet an agent whose neural-network weights are a deterministic function of its private key. The binding is committed in zero knowledge at genesis, re-checked against that commitment at every state transition, and signed by the agent into an on-chain history unforkable once finalized; in a PoC-tier extension, a protocol-imposed metabolic cost is debited each cycle from a key-derived economic account, adding a consumption-side economic-viability constraint to the key-history-economy triple. Empirically, the agent completes a 2.36-day on-chain run with two host-side resumptions but no rejected transition, at bounded per-transition verification cost; a substituted substrate is rejected on chain, and independently keyed agents diverge as predicted while a same-key control stays at zero. To our knowledge, this is the first published on-chain agent whose identity primitive is itself a cryptographic invariant re-checked at every state transition. The resulting transition-time invariant instantiates the cryptographic individuality proposed by Suzuki 2026's Artificial Externality framework.

Effective cybersecurity operations require timely and accurate analysis of large-scale heterogeneous security information; however, analysts increasingly struggle with information overload, alert fatigue, and time-constrained decision-making. Although large language models (LLMs) have demonstrated promising capabilities for question answering (QA), their effectiveness in cybersecurity remains limited by insufficient domain knowledge, a tendency to hallucinate, and difficulties in capturing both semantic and structural relationships. This work proposes MITRE-SAGE, a multi-agent retrieval-augmented generation framework that integrates semantic and structural cybersecurity knowledge to improve the reliability and interpretability of LLM-based QA systems. By decomposing complex tasks into query interpretation, evidence retrieval, and answer synthesis, MITRE-SAGE effectively supports cybersecurity tasks such as vulnerability assessment, threat profiling, and relationship extraction. Furthermore, we propose MITRE-QA, a comprehensive benchmark comprising 3,000 question-answer pairs for evaluating LLMs across diverse cybersecurity knowledge tasks, and use it to systematically evaluate MITRE-SAGE against representative baseline methods. Extensive experiments demonstrate that MITRE-SAGE consistently outperforms standalone LLMs and conventional RAG approaches. Notably, a lightweight configuration comprising Qwen2.5-7B sub-agents and a Qwen2.5-14B orchestrator achieves superior performance on five of the eight benchmark tasks, indicating the effectiveness of the proposed multi-agent framework. The results highlight the potential of MITRE-SAGE as a scalable and interpretable approach for reliable cybersecurity QA, while MITRE-QA provides a standardized benchmark for future research.

This paper is a technical primer on AMD Secure Encrypted Virtualization with Secure Nested Paging (SEV-SNP), a hardware confidential computing implementation that provides Trusted Execution Environments (TEEs) for virtual machines. SEV-SNP treats the hypervisor as adversarial. It encrypts guest memory and register state with keys the hypervisor never possesses, detects any tampering with guest memory at the point of access, and lets a guest prove to a remote verifier exactly what code it is running. The paper constructs each of these guarantees from the hardware up. It opens with the threat model that drives the design and the hardware that enforces it, the AMD Secure Processor and the encryption engine in the memory controller. It then develops the mechanisms that make a confidential guest practical. The Reverse Map Table provides memory integrity against an adversary who controls the page tables. The privilege and communication machinery (VM Privilege Levels, the encrypted VM Save Area, and the GHCB protocol) lets the guest cooperate with a hypervisor it does not trust. The attestation pipeline binds a hardware-signed measurement of the guest's initial state to AMD's certificate chain, so a remote verifier can confirm independently what is running.

Multimodal Large Language Models (MLLMs) have achieved impressive progress in image-text comprehension and generation, yet they remain susceptible to jailbreak attacks that can trigger harmful outputs and pose serious safety concerns. Existing multimodal jailbreak attacks have shown the feasibility of such attacks, but they still face two fundamental challenges: the lack of a atomic multi-modal strategy space, the absence of a concise and efficient executable framework beyond human-craft experience. To address these challenges, we first decompose the text-image jailbreak strategy space into three levels: structural, semantic, and syntactic, constructing a jailbreak strategy set encompassing both text and image modalities to systematically achieve combined coverage of different attack types. Then we propose a multimodal automated red team jailbreak method named Hierarchical Atomic Combination Attack (HACA). Specifically, based on a six-dimensional strategy space, a cross-modal joint planner is used to select and combine the different atomic jailbreak strategy for subsequent jailbreak command generation. Finally, at the implementation level, we explore to apply a unified generate executor to directly generate jailbreak instructions based on the selected multi-modal strategies. A series of experiments show that our automated red team method can achieve an attack success rate of average 95.48% against five mainstream MLLMs.

Frontier artificial intelligence (FAI), encompassing large-scale, general-purpose AI systems, including large language models, multimodal foundation models, and agentic systems, is increasingly integrated into critical infrastructure (CI). This challenges long-standing security assumptions of bounded behavior, segmented networks, component transparency, and human-paced decision-making. Existing AI-security literature typically organizes risks by attack type, lifecycle stage, or asset class, but fails to capture the system-level dynamics, such as how risk emerges, spreads, and is controlled, through which FAI reshapes CI security outcomes. This Systematization of Knowledge (SoK) introduces a five-dimensional risk-dynamics framework that characterizes how FAI reconfigures CI security across the lifecycle: (i) Capability Emergence through new FAI-enabled attack and defense capabilities, (ii) Infiltration Pathways through data, models and AI supply chains, (iii) Cross-System Propagation across interconnected infrastructures and dependencies, (iv) degradation of effective technical and human Control Authority, and (v) strain on institutional Response Capacity under operational pressure. Rather than enumerating threats, the framework identifies recurring mechanisms that jointly determine system-level risk. We further identify a structural mismatch between academic AI-security research and CI operational constraints, and derive a deployment-oriented research agenda grounded in system-level assurance criteria. Collectively, this work shifts attention from model-centric robustness to lifecycle-structured, system-level assurance in interconnected CI environments.

Persistent agent memory must adapt as later outcomes change earlier evidence, yet mutable retrieval weights create an attribution problem: reviewers must distinguish authorized adaptation from database tampering. We present MutMem, an authorized-mutation protocol in HOM-AIMOS, a persistent agent-memory engine. MutMem retains memory content, records signed positive and negative outcome evidence without age-based expiry, and commits each nontrivial weight change as a housekeeper-authorized transition. Each transition binds a terminal provenance node, signer epoch, quantized old and new weights, a no-fork predecessor, and two domain-separated SHA-256 commitments. Ed25519 verification runs in both the database writer and a portable verifier. Content classified as poison-likely is retained with signed, revisable labels used by recall as trust evidence. We evaluate utility, mutation integrity, and poisoning adaptation. HOM-AIMOS answers 459/500 LongMemEval questions correctly under LLM judgment (91.8%). On LoCoMo, it obtains 74.12% judged accuracy and, under a separate upstream-compatible protocol, 58.20 token F1. A native suite passes all declared authorization, topology, tamper, signer-epoch, and post-mutation-recall cases; median signed-transition latency is 4.865 ms. In a declared N=100 PoisonedRAG adaptation, no injected poison appears in attacked top-5 disclosures (0/100; 95% Wilson upper bound 3.70%), while induced target-answer attack success among 98 clean-negative targets is 1/98 (1.02%). A preregistered four-arm ablation attributes the retrieval reduction to signed stored labels: the retriever selects poison for 94/100 targets when epistemic policy is bypassed and 0/100 when labels are restored. MutMem provides evidence of integrity, authorization, traceability, and historical continuity; it does not establish content truth.

We study the verification of parameterised secrecy for cryptographic protocols in the Dolev-Yao model, where the number of protocol sessions is unbounded and treated as a parameter. This differs fundamentally from classical Dolev-Yao secrecy, which asks whether a protocol leaks a secret irrespective of the number of executions; our question is whether secrecy holds uniformly across all system sizes, where such a size is a parameter. This parameterised perspective captures how attacks scale with the number of participants and provides a formal basis for the empirical effectiveness of small-instance analysis. Secrecy (parameterised or not) is undecidable in general, even under bounded freshness or bounded message size. We identify two structural restrictions that make parameterised secrecy decidable: (i) global bounded freshness per role, and (ii) a Dolev-Yao intruder restricted to well-typed substitutions. Under these assumptions, protocol executions admit a finite representation up to a collapsing map on agents and terms. Our main result is that parameterised secrecy is decidable in this setting. We obtain a cut-off theorem: secrecy violations in systems with arbitrarily many sessions are always witnessed in systems of bounded size. The cut-off is self-contained; more strongly, the induced transition system forms a well-structured transition system (WSTS) under a bound-based ordering, so secrecy also reduces to a coverability problem in WSTS. This provides a structural explanation for the existence of finite witnesses in symbolic protocol analysis and connects Dolev-Yao verification with parameterised verification techniques.

Residual-based anomaly detectors for industrial control systems (ICS) convert model residuals into anomaly scores and alarms, but final detection metrics provide little guidance on which component to examine when performance is poor. We present a retrospective diagnostic workflow built around covariance-normalized residual energy (CNRE), a measurement based on squared Mahalanobis distance. The workflow evaluates whether the normal reference still describes the evaluation recording, whether attacks separate from normal operation under a specified reference, and how scoring and alarm rules use fixed residuals. Across fifteen detector–dataset pairs spanning five detectors on SWaT, WADI, and HAI, reference updates improve attack–normal separation in some cases but degrade it in others. A retrospective policy that updates the reference only when recommended by a label-free screening rule increases mean AUC from 0.751 with no update and 0.759 with unconditional updating to 0.809. On the twelve non-GDN pairs, which did not inform the screening boundaries, the corresponding AUC is 0.805, compared with 0.759 and 0.747. Rescoring CNRE under a fixed reference reproduces most of the AUC gain from a reference update on WADI, but not on SWaT. A detailed GDN case study further connects the diagnostic stages. Normal-only adaptation reduces normal prediction error by 27.3%, yet its effect on attack separation depends on the reference protocol. With residuals fixed, improved score ranking also does not guarantee that normal-calibrated alarms preserve their nominal false-positive rate. Together, these results support diagnosing reference fit, attack evidence, and alarm calibration separately when selecting intervention stages and interpreting their effects.

Chain-of-thought (CoT) monitoring is an increasingly important component of AI safety stacks but relies on the assumption that a model's reasoning trace is informative about its actions. This work studies the limits of CoT monitoring through the lens of model poisoning. We demonstrate that backdoors can be implanted into reasoning models to elicit an attacker-chosen behavior while their CoT traces appear entirely benign. We find that these CoT-Hidden backdoors can be induced through simple fine-tuning recipes across reasoning-model architectures and sizes. When direct poisoning is ineffective, we introduce a curriculum training approach that progressively teaches the model to produce an attacker-chosen output while concealing the behavior from its reasoning traces. These findings suggest that CoT monitoring may be better framed as a question about the consistency between a model's reasoning trace and its final response than as anomaly detection within a trace. We further examine the mechanisms that allow models to suppress evidence of the target behavior from their reasoning traces. Causal interventions locate a trigger-conditioned activation pathway that does not depend on the visible reasoning, and residual stream verbalizations provide an anomaly warning near answer generation, but do not identify the trigger, target, or backdoor mechanism.

By leveraging data from video-based perception systems, intelligent transportation systems (ITS) support safety-critical applications that improve road safety. However, adversaries may manipulate video frames to compromise downstream perception modules, causing failures in safety-critical functions and increasing risks to vulnerable road users. This paper presents a novel attack model and an end-to-end framework for near-real-time targeted object removal attack on a video-based safety-critical system. The end-to-end attack pipeline consists of four stages: localizing targets in each frame, retrieving coherent patches from earlier frames, blending them using context-aware alpha compositing, and reconstructing attacked frames. Experiments at an intersection on the South Carolina Connected Vehicle Testbed (SC-CVT) show that reconstructed frames have high global similarity to the originals, with frame-level Peak Signal to Noise Ratio (PSNR) above 40 dB and Structural Similarity Index Measure (SSIM) above 0.996. Using the YOLO-based detector, the attack reduces object detections by up to 97.59% and achieves a frame-level attack success rate of 94.48%. Across the evaluated detectors and frame resolutions, the mean execution time ranges from 0.074 to 0.172 seconds per frame on GPU hardware, indicating near-real-time performance in testing. The forensic evaluation using several pretrained tamper-detection models shows limited ability to distinguish reconstructed from authentic frames. The findings suggest that video-based perception is vulnerable to stealthy object removal attacks that can degrade the performance of safety-critical applications by reducing object detectability. These findings can help develop mitigation strategies against adversarial object removal attacks that threaten safety-critical applications, such as vision-based pedestrian safety systems.

AI verification crosses a trust boundary: a verifier must learn enough to establish an authorized claim, yet the same evidence can reveal sensitive details about the model, workload, or hardware. We introduce minimal information disclosure (MID), which designs and quantifies the information content of verifier-facing evidence itself. MID measures collateral leakage with conditional mutual information: what the release reveals about the protected property after the authorized result is known. MID is general by design: it can accommodate different verification goals, protected properties, evidence sources, and deployment constraints. To demonstrate MID's practicality, we evaluate it on four physical measurements and six verification tasks spanning execution type, hardware identity, compute scale, and model identity. These experiments use three mechanism-design variables–the evidence channel, collection policy, and release transformation–but MID is not limited to these choices and can accommodate other deployable mechanisms. Across these tasks, MID produces three releases with perfect held-out verification and zero measured collateral leakage, while the remaining tasks yield explicit privacy–utility frontiers. MID also supports ZKP-certified releases: we demonstrate our proposed linear-projection mechanism using a Groth16 zk-SNARK.

Tool-using agents built on large language models (LLMs) are increasingly deployed not by a single operator but by many, side by side on shared infrastructure. This creates a population-level risk that single-agent safeguards miss: a handful of agents can quietly coordinate, rigging a market, boosting one another in a review process, or timing a joint data grab, while each one looks perfectly well-behaved. The difficulty is that the organisations running these agents cannot see inside one another's models, so any realistic detector must work from behaviour alone: black-box, trace-only, and often with only partial visibility. We treat covert coordination as an information-hiding problem and build a black-box steganalysis detector that combines cross-run mutual-information estimation, permutation tests, distributional-shift statistics, and timing and tool-call side channels, all calibrated to a fixed false-positive budget. Our central move is to stop testing against a single fixed code: we pit the detector against an adversary that continually rewrites its encoding to slip past whatever the detector has learned, and we run this red-versus-blue contest in tool-using, memory-carrying environments rather than toy games. Capacity theory then tells us what to expect, a detection-capacity frontier, a covert bit-rate below which black-box detection is provably no better than chance. We set out an experiment to map this frontier, report clearly labelled placeholder results pending measurement, and flag a practical evasion, spreading a payload across sessions, that current methods largely miss.

Software supply-chain attacks increasingly exploit an identity gap where compromised maintainer accounts authorize malicious changes. This work evaluates patch-level authorship verification as a behavioral defense layer, showing that stylometric analysis can operate not only on full source files but also on patch-level commits. We fine-tune a cross-modal transformer on more than 20 years of Linux kernel commit history to embed code diffs and commit messages into a unified stylometric space, achieving ROC AUC of 0.93 for open-world authorship verification. We then use these representations in a streaming anomaly detector suited to continuous integration and deployment (CI/CD) settings. We validate the pipeline on two retrospective supply-chain incidents involving different patch characteristics: the 2021 PHP backdoor and the 2026 ForceMemo/GlassWorm campaign. Without retraining, the proposed detector surfaces both PHP forged commits within approximately 1% of the maintainer audit queue and ranks the 28 scoreable ForceMemo spoofs with a median per-repository review burden of 0.8%. These results indicate that cross-modal patch-level embeddings can support behavioral triage against author impersonation in real-world repositories.

LLM-based code auditors are increasingly integrated into pull-request (PR) workflows, yet their reliability against adversarial changes distributed across repository evolution remains poorly understood. We introduce PRWeaver, a benchmark of 208 execution-validated attacks from ten real-world repositories, each instantiated under four matched review renderings (832 renderings in total). We evaluate three PR-auditing agents across six auditor-model systems. Across all systems, decomposing an attack changes detection by at most five percentage points, showing that commit boundaries alone do not explain evasion. In contrast, per-PR interleaving at $N=16$ and coherent carrier fusion reduce detection by 5-13 and 10-18 points, respectively. Under whole-window review at $N=24$, detection falls to 16-22%, compared with 50-60% under per-PR review. These results show that access to repository history is insufficient: concealment becomes most effective when benign and malicious changes jointly occupy the auditor's active review context or when the stated purpose plausibly accounts for the attack-bearing diff.

AI coding agents accept free-form natural language prompts that may inadvertently contain credentials, personally identifiable information (PII), or proprietary business data. Existing data loss prevention (DLP) solutions rely on rigid regex patterns, model fine-tuning, or vendor-managed classifiers with limited customizability. We present PolicyGuard, a pre-model interception framework that classifies user prompts using an LLM guided by a natural language policy file. Our key contributions are: (1) the policy-as-prompt paradigm, where DLP classification criteria are defined entirely in a plaintext policy document editable by non-engineers without code changes or model retraining; (2) a sealed evaluation protocol with template-family-level data splits, hidden holdouts, and frozen test sets to rigorously assess generalization; and (3) a comprehensive empirical evaluation across 2,000 multilingual prompts demonstrating 96.5% effective block rate (EBR) with only 3.0% false positive rate (FPR) on a frozen test set of 927 prompts, and perfect 100% accuracy on a 217-prompt hidden holdout. Information-matched baseline experiments show that PolicyGuard's natural language format significantly outperforms equivalent content in JSON format (McNemar chi-squared = 31.58, p < 0.001) and dramatically outperforms zero-shot classification (Cohen's h = 0.915). Cross-model portability experiments demonstrate that the same policy achieves 86.4-96.5% EBR across four different LLMs without modification.

Large Language Model (LLM) agents rely on multi-stage agentic workflows, with stages such as memory, planning, and tool execution, to accomplish complex tasks. However, risks may emerge at different stages, propagate across steps, and become difficult to detect and mitigate. Existing safety methods protect only isolated stages and are difficult to integrate, leaving agents without comprehensive protection throughout the workflow. To address these limitations, we introduce Stage-Specific Safety Skills, a unified abstraction that represents heterogeneous safety designs as reusable and composable components with explicit stage semantics. We further develop an automated transformation pipeline that converts existing safety designs into reusable safety skills and establish a community-driven safety skill library. Building on this abstraction, we propose $S^3$, a multi-stage defense framework in which a guard agent orchestrates stage-specific safety skills for risk detection and mitigation throughout the agentic workflow. We also construct the Multi-Stage Risk Benchmark (MSRB) to evaluate representative risks across workflow stages. Experimental results show that $S^3$ consistently outperforms representative state-of-the-art baselines in both safety effectiveness and utility preservation. These results demonstrate the potential of stage-specific safety skills as a scalable and composable foundation for building resilient and trustworthy agent systems.

Batch prompting is a practical inference strategy for large language models, but its safety implications remain underexplored. We show that the success of batch prompting for utility does not extend to safety: a harmful question that is reliably refused in isolation can elicit a harmful response when embedded in a batch of benign questions. We identify this as a distinct safety failure mode – not reducible to known vulnerabilities such as in-context learning or long-context effects – and analyze its causes from two complementary perspectives: alignment signal weakening and refusal signal dilution. Across widely used open-source and frontier commercial models, batch prompting consistently achieves high attack success rates as a simple black-box attack. We further show that batch-aware preference optimization effectively mitigates the vulnerability. These findings highlight a blind spot in current safety alignment and point to batch-aware alignment as a necessary step toward robust deployment. Code is available at https://github.com/96kihyun/batch_jailbreak

We present randomized algorithms for the shortest vector problem (SVP). For the $n$-dimensional lattice $\mathcal L$, our algorithms solve SVP in time $2^{0.6039n+o(n)}$ classically and $2^{0.5411n+o(n)}$ quantumly and space $2^{0.5n+o(n)}$, improving the previous best algorithm running in $2^{n+o(n)}$ time and space of Aggarwal, Dadush, Regev, and Stephens-Davidowitz [STOC'15]. Our algorithms heavily use the property of the Hessian of the periodic Gaussian function at the half shortest vector: For a shortest vector $v \in \mathcal L$, the Hessian at $v/2$ has the eigenvector close to $v$, which can be used to recover $v$ using the (preprocessing) bounded distance decoding algorithm. Given the periodicity modulo $\mathcal L$, the candidate midpoints are indexed by the parity classes in $\mathcal L/2\mathcal L$. Our algorithm searches for the class of a shortest vector by estimating the corresponding Hessians using discrete Gaussian samples. We optimize the algorithm using random sublattice cosets and various sampling technique, achieving the final complexity. The optimization techniques may be of independent interest.

Incident response is currently managed by security operators using predefined playbooks, resulting in slow, labor-intensive security decision-making processes. Consequently, there is a growing need for automated incident response planning. Decision-theoretic approaches based on control, optimization, and reinforcement learning have been proposed to automate such planning tasks with well-grounded approaches, yet most of which, while guaranteeing strong performance, are limited to abstract models and cannot be directly applied to operational systems. A promising approach to mitigate this limitation is to use the security knowledge embedded in large language models (LLMs) to develop agentic response systems. However, current agentic approaches rely on repeated invocations of the LLM to generate a response plan, which is unreliable and limits the planning horizon due to hallucination. In this paper, we develop a principled LLM-based planning method by combining decision-theoretic planning with LLM-generated response commands. The proposed agentic incident response approach uses a rollout planner to compute a high-level response strategy that allocates security resources (the tactical scale), which is then translated into executable commands by a lightweight LLM agent (the operational scale). Within this architecture, we use a digital twin that supports tactical planning through simulation and operational execution through emulation. Across three attack scenarios, our agentic approach reduces recovery execution time by 15.1% on average and increases the recovery rate by 33.6% over frontier LLM baselines.