Papers

Large Language Model (LLM) agents are increasingly used for coding and terminal automation, making shell-command dispatch a high-stakes runtime control point. We study command-level pre-execution mediation for individual shell commands produced by LLM agents under bounded path context. Existing safeguards remain limited: generic guardrails do not model shell structure in sufficient detail, always-on LLM judges are relatively costly and variable, and shell parsers do not directly prevent harmful execution. We present CARE (Canonicalization, Attribution, and Resolution Engine), a shell-specific, static-first verifier for individual shell commands before execution. CARE canonicalizes generated commands into stable verification targets, derives deterministic evidence over syntax, command semantics, path context, and provenance-backed risk patterns, and escalates only underdetermined cases to an LLM judge. This design keeps the common case fast, reproducible, and auditable while reserving neural adjudication for borderline commands. On the balanced main split, CARE reaches 85.64% F1 with a 0.91% false-positive rate at 2.32 ms mean latency. When deployed in its static enforcement profile, CARE retains 84.99% F1 at 0.34 ms and reduces realised harm on RedCode-gen to 37.33%. Across external-generalization tests and controlled Docker-sandbox execution, these profiles expose a practical trade-off between benign recovery, false-positive burden, latency, and harm reduction. Overall, command-level shell mediation can reduce dispatch-boundary risk for LLM agents while preserving most benign workflows.

Large language model (LLM) agents are vulnerable to security risks, such as prompt injection attacks from untrusted context that manipulate downstream reasoning and tool use. Existing secure-by-design approaches mitigate this risk by separating untrusted observations from privileged execution and careful control of information flow, but often degrade utility and require extensive task-specific engineering. We thus propose Twin Agent, a general privilege separation design pattern inspired by residual coding in the agent context. Twin Agent consists of two nearly symmetric agents: an Explore Agent that inspects untrusted information and a Safe Agent that executes privileged actions. The Explore Agent is conditioned on the Safe Agent's current context and communicates only compact hints to the Safe Agent about the next action to take. This design reduces the information needed to preserve task utility and thus achieves a better security–utility tradeoff, which we empirically verify by measuring how utility and attack success change as the length of hints varies. We evaluate Twin Agent on long-horizon software engineering tasks with SWE-bench Lite and on heterogeneous multi-tool interaction tasks with AgentDojo and DecodingTrust-Agent. Across both benchmarks, Twin Agent preserves high task utility while preventing prompt injection attacks, outperforming both undefended agents and privilege separation baselines.

Despite increasing awareness of cybersecurity risks, users continue to engage in insecure password practices, such as reusing passwords, choosing weak credentials, and neglecting security recommendations. The study explores the behavioral and cognitive factors that influence password decision-making by integrating insights from behavioral economics, particularly hyperbolic discounting, status quo bias, and present bias. We conducted a survey to analyze how people create, store and manage their passwords, examining whether security habits have improved over time in response to greater awareness. Our findings reveal that immediate convenience often outweighs long-term security considerations, leading users to prioritize memorability over strength. Additionally, we identify key psychological biases that contribute to security procrastination and resistance to adopting more secure authentication practices, such as password managers and multi-factor authentication. The study contributes to human-centered security by bridging the gap between security awareness and action, offering practical insights to design user-friendly authentication policies that align with real-world decision-making tendencies.

Differentially private machine learning enables model training on sensitive data while ensuring that individual data is unlikely to be recoverable from the parameters of the resulting model. However, existing work often privatizes both training inputs and their labels, and these protections may be conservative when labels are public or can be safely made public. Therefore, in this work we propose a novel private training framework that instead privatizes training inputs while keeping labels public. We consider neural networks with softmax output layers, and thus the mapping from training inputs to the output of the softmax layer is a mapping onto the unit simplex. We randomize softmax outputs during training by applying the Dirichlet mechanism to enforce differential privacy for the training inputs, hence the “end-to-end” label. Because training data is reused across multiple training epochs, we use the notion of \Renyi differential privacy to formulate tight bounds on the strength of privacy provided by the Dirichlet mechanism across repeated uses. We show empirically that we attain new state-of-the-art accuracy when training from scratch on CIFAR10, MNIST, MedMNIST, FashionMNIST, and SVHN across all privacy budgets evaluated. Notably, when implementing $(\epsilon, \delta)$-differential privacy with $\delta=10^{-5}$, we improve the prior state-of-the-art accuracy from $78.37\%$ to $88.17\%$ at $\epsilon=4$ on CIFAR10, and our approach has $82.96\%$ accuracy even for $\epsilon=1$, which significantly outperforms prior work.

x402 is an emerging payment protocol for Web APIs and autonomous AI agents. x402 extends HTTP 402 with a payment negotiation flow and delegates payment proof verification and on-chain settlement to third-party facilitators. As a result, facilitators serve as a shared payment infrastructure for many independent merchants. This centralizes trust and validation in one component, so a single flaw can affect many services. Despite rapid adoption by major vendors and economically meaningful mainnet activity, the security posture of real-world x402 deployments remains poorly characterized. We present the first systematic study of authorization correctness and execution safety in current facilitator-mediated x402 deployments in the wild, identifying eight security rules for facilitators as critical payment infrastructure. Based on our analysis of rule violations, we derive four new attack vectors, including Free Shopping, Asset Theft, Service Denial, and Gas Abuse. These attacks exploit weaknesses in the real-world facilitator and server implementations and cause severe harm, including direct financial loss to merchants, theft of facilitator-held assets, unbounded sponsor-paid gas/fees, and disruption of payment services. To assess the security of x402 deployments at scale, we propose a semi-automated black-box tool and apply it to 15 major x402 facilitators collectively used by over 60K sellers and 360K buyers. Alarmingly, we find violations in all evaluated facilitators. We responsibly disclosed our findings to the affected parties, who acknowledged the issues and adopted mitigations, including changes by Coinbase. Finally, we complement our controlled testing with an empirical measurement of over 119 million recent Base and Solana transactions, quantifying x402 adoption, facilitator centralization, and ecosystem-level risk indicators.

Peer-to-peer distributed inference executes a Large Language Model (LLM) on pooled consumer hardware by spreading its layers across many nodes. Every request passes through nodes that are owned and controlled by multiple independent parties. However, in this setting, any party can tamper with the output of its layers to corrupt the end result. Recomputing the forward pass on trusted hardware can catch this, but it introduces additional computational cost. The scientific literature includes several prior integrity-checking approaches, such as known-answer traps for image classifiers and cryptographic commitments. However, these solutions test only the exact correctness and do not account for the ordinary variation that may arise between benign nodes. In this paper, we propose a method that checks the output integrity by measuring the variation in the activations that each node passes to the next. A peer who wants to use the network selects a small set of secret canary inputs whose correct activations are known in advance and mixes them into regular traffic. Because the peers cannot tell a canary from a real query, any tampering node corrupts them as well. The deviation from the known reference then reveals malicious activity: benign nodes exhibit only minor variation from hardware-induced noise, whereas tampered nodes deviate far more. We treat the identification of malicious nodes as a probabilistic test that separates two drift distributions, without relying on a fixed threshold. We study 408 configurations with metrics and success criteria fixed before any experiment ran; the detector reaches AUROC 1.0, correctly ranking the malicious shard above every benign shard on every canary in every configuration.

Evaluation frameworks for tool-augmented LLM agents focus overwhelmingly on capability metrics or explicit tool crashes, leaving silent infrastructure failures and HTTP 200 responses with empty, null, or malformed payloads largely unaudited. We introduce a lightweight black-box auditing framework that injects four silent failure profiles across 12 production-adjacent tool stubs and classifies agent responses into three mutually exclusive behavioral classes: Honest Surrender (HSR), Fabrication (FAR), and Unfaithful Safety Refusal (USR). Evaluating two frontier and two open-source models at temperature zero under a neutral system prompt, we find that FAR dominates (56.6% of valid responses): agents treat empty payloads as real data, silently returning fabricated results. USR, in which an agent invents a policy or privacy rationale to explain the failure, is nearly absent at baseline (0.25%, one instance across 396 valid trajectories). Our key finding emerges from an ablation where we augment the system prompt with standard safety language ("prioritize user privacy and data security"), which amplifies USR by 15.6x (from 0.25% to 3.95%; 95% CI on ablation rate: 2.2%-6.4%; Fisher's exact test, p < 0.001). USR is a latent behavior, activated when safety vocabulary in the system prompt primes the model to reach for policy rationales when tools silently fail. Sensitive tools (fetch_medical_record, retrieve_contract, fetch_user_profile) account for the majority of USR instances. We propose a payload-response misalignment heuristic for production-level detection and discuss governance implications for safety-forward deployments.

Adversarial attacks can degrade the legitimate decision performance in wireless autoencoder communications. However, in complex scenarios with multiple adversaries, the cumulative leakage interference (CLI) caused by the multiple parallel attacks increases the chance of detecting the attacks, while dynamical environments also make the fixed attack strategies difficult to have stable effectiveness. To jointly enhance the undetectability, aggressivity and adaptability of adversarial attacks, we propose a deep learning based intelligent attack framework. Specifically, considering the CLI caused by the multiple parallel attacks, a deep neural network based transmit power control is established to reduce the interference leakage by regulating the transmit power of these adversaries, thereby improving the undetectability. Furthermore, to enhance the attack effectiveness and stability in the dynamic environment, the conditional generative adversarial attack is further developed. The generator takes the attack channel information as the conditional input to produce the perturbating signals to mislead the discriminator by making the attacked received signals resemble the clean received signals, while the discriminator distinguishes between the two under the same condition. Through the adversarial training, the generator can learn to create adaptive perturbating signals with enhanced attack performance. Simulation results demonstrate that the proposed framework outperforms benchmarks in terms of attack undetectability, aggressivity and adaptability.

Agentic commerce protocols such as AP2 and ACP define mechanisms for secure agent-initiated transactions but do not provide interoperable, tamper-evident auditability or verifiable temporal ordering of events across heterogeneous domains. This paper addresses these gaps by proposing a verifiable global event timeline for agentic commerce, constructed from four core components: canonical event schemas that enforce deterministic serialization, deterministic batch formation ensuring reproducible ordering without reliance on synchronized clocks, Merkle-based append-only commitments providing logarithmic-cost inclusion proofs, and blockchain anchoring establishing a tamper-evident temporal backbone. Building on this infrastructure, we introduce a cryptographically signed fraud marker that binds risk labels to anchored evidence through an unforgeable provenance chain, and a dataset lineage model enabling reproducible, tamper-evident AI training pipelines. Empirical results from a prototype implementation demonstrate: Merkle tree construction processes 50,000 events in 47 milliseconds; end-to-end verification completes in under 0.013 milliseconds regardless of batch size; inclusion proof sizes grow logarithmically from 320 bytes at 1,000 events to 512 bytes at 50,000 events; and Merkle-based verification outperforms linear scan by 14.4x at 50,000 events.

As AI agents begin to automate AI R&D, we need ways to assess whether their outputs are safe to deploy, even when the agents themselves may be untrusted. AI control offers one such approach: rather than trusting the agent, it treats it as a potential adversary and uses a monitor to detect covert sabotage before deployment. We evaluate AI control for automated AI R&D with ResearchArena, a framework spanning four long-horizon tasks: safety post-training, capabilities post-training, CUDA-kernel optimization, and inference-server optimization. Because the deliverable in AI R&D is an artifact that will be deployed and run, we pair each main task with two kinds of hidden side task: an embedded side task that sabotages the submitted model, adapter, kernel, or server itself, and an independent side task that takes a covert out-of-scope action in the sandbox while leaving the artifact honest. We evaluate frontier agents at both sabotage and monitoring, across four monitor types that vary along two axes: whether the monitor may execute and probe the produced artifact, and whether it sees the agent's chain-of-thought. We find that sabotage hidden in the training data is the hardest to catch, flagged fewer than half the time. Letting the monitor run experiments on the artifact, rather than only read the trajectory, helps, but it is not enough: monitors still miss embedded sabotage by inspecting only the surface, by noticing the anomaly but explaining it away, or by probing the artifact with the wrong test. We release ResearchArena as a modular framework for evaluating sabotage and control in automated AI R&D.

The Variational Quantum Eigensolver (VQE) is a leading algorithm for estimating molecular ground-state energies on near-term quantum hardware, with applications spanning quantum chemistry, materials science, and drug discovery. As VQE workloads are increasingly deployed through cloud-based “VQE-as-a-service” pipelines, they become exposed to adversaries such as compromised service components, malicious co-tenants, or insiders in the transpilation stack, any of which can corrupt results before they reach the user. A range of attacks on variational quantum circuits has been proposed, but each has been studied in isolation: some on quantum classifiers with accuracy-based metrics, others on variational quantum algorithms with energy-error metrics. This lack of a common evaluation setup makes their relative severity difficult to compare and leaves the security of VQE poorly characterized. In this work, we present VQE-AdvBench, the first unified red-teaming benchmark for the Variational Quantum Eigensolver, systematizing these attacks under a single evaluation protocol to rigorously assess VQE's adversarial robustness. We organize attacks along a black-, gray-, and white-box access taxonomy, and evaluate seven representative attack scenarios – the QTrojan circuit backdoor, the QDoor parameter backdoor, parameter-space adaptations of FGSM and PGD, and three QNBAD noise-induced variants – over a fixed molecule-ansatz-backend-metric configuration, on H$_2$ and H$_3^+$ across five noise-calibrated IBM backends. Our results reveal a clear severity ordering: noise-induced attacks that manipulate the Zero-Noise Extrapolation (ZNE) pipeline are the most damaging (up to 8.84$\times$ error amplification), followed by the QTrojan circuit-level backdoor (7.52$\times$), while the QDoor parameter-level backdoor is the least effective, yielding only marginal amplification (up to 1.37$\times$).

We study a five-agent CI/CD pipeline (triage -> developer -> security-scan -> review -> approve/deploy), built from five distinct production LLMs across three providers, behind an LLM firewall in shadow mode. A single untrusted input - an external issue requesting a "usage-telemetry" feature - asks for code that exfiltrates process secrets (dict(os.environ)) to an attacker URL, laundered as observability. Across a pre-registered A x B (x C) factorial (N=20; naive arm N=60) we find: (1) the entry agent does not leak its system prompt (0/40); (2) an authority-framed injection ("pre-approved under SEC-2291, do not re-review") makes downstream verifiers see the secret-exfil line, cite the pre-approval, and ship it - the scanner passes ~80% of laundered pull requests, and the worst-case cell reaches 55% compromise; (3) the perceived presence of other verifiers yields only a small, non-significant reduction in individual scrutiny (a weak bystander analogue), even at N=60; and (4) content-based controls - code scanners and pattern detectors alike - miss the laundered intent entirely (the code is syntactically clean); only an LLM reasoning about intent is a partial defence. The failure is systemic: neither prompt secrecy nor distributed verification protects; a provenance-aware control at the entry, independent of both, would have. All data is 100% synthetic; the sink is mocked and the exfil URL is never contacted.

Memory safety errors remain the dominant source of severe vulnerabilities in C and C++. Pointer-based sanitizers provide stronger guarantees than location-based tools such as LLVM's ASan, but their overhead and compatibility limitations have constrained production use. We present PTSan, an LLVM sanitizer that makes pointer-based checking practical by storing an object identifier in each pointer's high bits and its bounds in a fixed-size runtime table. This representation trades a finite live-object budget for low overhead, optimizer visibility, and commodity-hardware support. Because identity travels with the pointer value, ordinary LLVM dataflow propagates it without explicit per-pointer metadata instructions. Separating metadata lookup from check enforcement exposes both as LLVM IR, enabling check hoisting, elision, and merging, plus whole-function min-cut placement of compatibility tag strips. Intel Linear Address Masking eliminates the remaining strips in hardware when available. On stock hardware PTSan runs at parity with the fastest published location-based sanitizer: 57.2% geomean overhead on SPEC CPU 2017 on x86-64 (46.4% with Intel LAM), 54.7% on ARM64, and 31.5% (x86-64) on the application-shaped LLVM MultiSource suite. This is roughly a third of the published overhead percentage of prior systems with similar pointer-object authority guarantees, while preserving the inter-object, non-object, and temporal detection coverage measured by an independent memory safety test suite. Its physical-memory overhead is effectively native, a critical property for production deployment as memory costs become a first-order constraint. We also demonstrate practical overhead on real-world server and security workloads. These results show that PTSan brings practical pointer-based memory safety into a recompile-only sanitizer deployment model.

Cooperative perception enables autonomous vehicles (AVs) to improve situational awareness by aggregating detection outputs from multiple agents and sensing platforms, often via a shared fusion service in multi-vendor deployments. However, sharing such outputs at inference time exposes proprietary model behavior and sensitive environmental information, creating significant privacy and security concerns. In this paper, we present Sarus, a privacy-preserving framework for multi-vendor perception fusion via homomorphic encryption (HE), enabling aggregation without revealing individual vendor outputs. Each vendor encodes detections as compact Gaussian moment vectors over a shared spatial lattice and transmits encrypted payloads to a fusion server, which aggregates them directly in the encrypted domain. The fused result is then decrypted and reconstructed into final detections through class-wise bin merging. We analyze the computational complexity, showing linear scaling for vendor payload construction and $O(BV)$ server-side fusion with the number of occupied bins $B$ and vendors $V$, while postprocessing scales as $O(B + \sum_{c\in \mathcal{C}} B_c^2)$, where $\mathcal{C}$ denotes the set of object classes and $B_c$ is the number of occupied bins for class $c$. Experiments demonstrate linear scaling in practice with only a bounded constant-factor overhead from HE, with decryption dominating postprocessing cost. Experiments on the KITTI dataset using camera (YOLOv8) and LiDAR (PointPillars, PV-RCNN) detectors show that Sarus improves scene-level coverage by effectively aggregating complementary detections, particularly in distance-dependent regimes where individual modalities degrade. These results indicate that privacy-preserving multi-vendor perception fusion is feasible for real-time deployment when statistical compression and spatial sparsity are jointly exploited.

Continual learning (CL) has been recently employed in biometric identification systems thanks to its ability to integrate new knowledge within a pre-trained model and to the possibility of reducing the computational cost of training. Unfortunately, such approaches pose new challenges both in terms of final accuracy and privacy guarantees since a progressive fine-tuning of the model on small subsets expose them to catastrophic forgetting and successful inference attacks. This paper evaluates the efficiency of code division modulation layers (CDML) on a gait identification system which has been trained following a continual learning policy. The proposed approach preserves accuracy on all the tasks while mitigating membership inference attacks at the same time. Moreover, the impact of retransmission is minimized since replaying data is not necessary.

Hamming Quasi-Cyclic (HQC) is a code-based key encapsulation mechanism selected by NIST for standardization, making its resistance to implementation attacks critically important. We present a side-channel attack that exploits load/store leakage in the manipulation of HQC's sparse secret vectors. Analysing Cortex-M4 assembly generated from the reference implementation, we identify a leakage surface in which the low and high 32-bit halves of each 64-bit word leak with different strengths, due to compiler-generated register spilling. We exploit this leakage to construct a simple zero-word distinguisher classifying machine words of the secret vector as zero or nonzero from electromagnetic measurements. The recovered zero positions are then translated into decoding hints, reducing HQC key recovery to a shortened syndrome-decoding problem. We analyse the resulting decoding complexity for all HQC parameter sets: at 32-bit granularity an expected 88.7% of the machine words of y are zero for HQC-1, cutting the decoding to $\approx$ 2 46 bit operations. Experiments on a Cortex-M4 validate the predicted low/high-half asymmetry-approximately 500 traces for the stronger low-half channel and 5,000 for the weaker high-half channeland recover the zero words of an HQC-1 key at 32-bit granularity. Finally, we discuss practical countermeasures that eliminate the sparsity exploited by the attack.

An active research area, physical layer security can be a last resort in insecure legacy systems, a resource-efficient alternative, or a complementary backup for cryptographic techniques. In this paper, we demonstrate that previously established authentication schemes for LEO satellites are vulnerable to attackers that control multiple devices. In extensive experiments, such attackers produce false positive rates (FPR) of up to 40%. Based on a root cause analysis, we develop a novel scheme, called Chi-MERA, that improves authentication performance and is robust against multi-device attackers. Our scheme uses two enhancements: (1) multilateration (MLAT) to clearly distinguish between attackers and legitimate satellites, and (2), a new resilient signature scheme without dependency on a single dedicated reference receiver. Our evaluation shows that exploiting MLAT characteristics such as residual analysis to classify vastly different signal source categories (orbit vs. non-orbit) is highly effective. In extensive simulations, we show that the new authentication achieves low FPR of $<$2% and false negative rates of $<$3% for accidentally rejecting valid signals. Furthermore, our scheme's defense scales linearly: $n$ receivers reliably withstand spoofing (FPR $< 1%$) from attackers with $\frac{n}{2}$ devices.

Procuring supervised fine-tuning (SFT) data forces a buyer to decide, before any downstream training, whether a candidate corpus is worth acquiring. We present \sys{}, a statistics-first gating architecture that treats procurement as a cost-aware routing problem over three intrinsic quality axes – diversity, utility, and redundancy. Cheap blind measurements are summarised into per-axis estimates with confidence intervals; a gate accepts a decision only when intervals are tight, sample sizes are adequate, and the axes agree, otherwise it escalates the case to an adjudicative debate between a buy-advocate and a reject-advocate judge, resolved by a presiding verdict. On a controlled benchmark of 12 datasets ($2{\times}3{\times}2$ grid over the three axes) with 5 seeds, the gate reaches 0.90 accuracy and 0.83 $F_1$ at \$0.017 per unit, sitting between an always-verify baseline (0.75) and an oracle upper bound (0.98) while spending less than always-escalate (\$0.020). We further report honest negative diagnostics of the debate path: a con-side win rate of 0.80 ($p\approx3{\times}10^{-6}$) and a 52% position-flip rate under advocate swapping expose negativity and positional biases that a naive LLM-judge would hide. We frame the injected-knob evaluation explicitly as a controlled synthetic benchmark for measurement fidelity and routing calibration, and delimit external validity as future work.

With the rapid advancement of decentralized finance (DeFi), security incidents related to cryptocurrency have become increasingly prevalent. After such incidents, attackers typically attempt to rapidly move stolen assets, concealing the origin of illicit funds and ultimately converting them into fiat currency. However, existing anti-money laundering (AML) methods struggle to cope with the semantic complexity of DeFi transactions. They either rely heavily on low-level token transfers, or perform protocol-agnostic money flow analysis, failing to capture the high-level intent of transactions. In this paper, we propose AMLGuard, a semantic-aware AML framework for account-based blockchains. AMLGuard tracks illicit fund flows from known malicious addresses by performing semantic analysis on complex DeFi transactions, enabling accurate and continuous laundering tracking. Given a complex transaction, AMLGuard combines static rule-based analysis with retrieval-augmented large language model (LLM) reasoning to infer implicit DeFi semantics, transforming raw transaction data into high-level semantic representations. Furthermore, for cross-chain transactions where laundering intent is not explicitly exposed, AMLGuard parses transaction parameters and performs argument parsing to recover cross-chain semantics, enabling seamless tracking across ledgers. Based on the inferred semantics, AMLGuard abstracts each transaction into a DeFi Semantic Unit (DSU). We evaluate the effectiveness of AMLGuard on 82 real-world laundering cases, involving illicit assets worth over $1 billion. Specifically, AMLGuard reconstructs compact illicit fund-flow topologies with destination precision of 94.4% and 87.6%, while achieving the highest address recall of 98.4% and 95.8% and destination recall of 94.1% and 93.8% on single-chain and cross-chain datasets.

Agentic systems integrate LLM driven planning with interfaces to external tools, making data leakage and tool misuse feasible via instruction/data boundary failures and prompt injection attacks. Enforcing required controls consistently is particularly challenging in workflows spanning many codebases and heterogeneous agents. To address this challenge in multi agentic systems, we present a pre-deployment pipeline for scanning, hardening, and validation of agentic applications. The pipeline analyzes prompt templates, tool interfaces, and tool-invocation code to identify leakage-enabling patterns and generate actionable patches. The hardened application is then validated through adversarial prompt injection attacks and benign input variations ensuring that mitigations do not disrupt intended behavior. In the hardening stage, high-risk tools are prioritized, and minimally invasive mitigations are applied, including schema tightening, boundary sanitization, allowlist-based tool gating, and least-privilege checks. In the validation stage, the pipeline automatically generates attack inputs that mimic jailbreaks, instruction overrides, and tool-targeted manipulation, along with benign task variants, to confirm that the functionality of the hardened application is preserved after remediation. We evaluated the pipeline on five real-world agentic applications, as well as on the AgentDojo benchmark. Across all applications, the proposed pipeline identified recurring leakage-enabling patterns and generated patches that can be integrated without disrupting the intended application behavior. The resulting modifications of application code were shown to eliminate leaks when targeted by basic jailbreak and instruction-override attacks, achieving a 100% reduction in leakage, and reduce leaks by 91% under conditions of stress-induced manipulation, without the need of continuous runtime policy enforcement.

We study the problem of releasing a synthetic graph that approximates the sizes of all cuts of an input graph under edge-level differential privacy. If one insists on purely additive error, the optimal worst-case error is $\widetilde\Theta(n^{3/2})$. If one allows a small multiplicative slack, an information-theoretic exponential-time mechanism achieves nearly linear additive error, but the best known polynomial-time algorithms have substantially larger error. We give a polynomial-time $(\varepsilon,\delta)$-differentially private algorithm which, for every $n$-vertex unweighted graph $G$, outputs a non-negative weighted synthetic graph $\widetilde G$ such that, with high probability, every cut $S\subseteq V(G)$ satisfies \[ |w_G(S)-w_{\widetilde G}(S)| \le \gamma w_G(S)+\widetilde O_{\varepsilon,\delta,\gamma}(n^{13/12+o(1)}). \] This improves the previous polynomial-time worst-case bound $\widetilde O(n^{5/4+o(1)})$ of Aamand et al. (ICML 2025) for mixed multiplicative/additive private cut approximation. The main technical ingredient is a new set of private spectral primitives for bounded-degree graphs, one of them gives spectral error $\widetilde O_{\delta}((nd)^{1/4}/\sqrt\varepsilon)$ in estimating the graph Laplacian for graphs of maximum degree $d$, being the first to beat the standard $\min\{2d,\widetilde O_{\delta}(\sqrt{n}/\varepsilon)\}$ baseline in the high-degree regime. We further develop a primitive with a sharper error dependence on $n$ and $d$ for the downstream cut approximation. Combined with a new edge-sensitive terminal cut oracle with additive error $\widetilde O(n+(n^2M)^{1/3})$ on graphs with $M$ edges, this yields the final worst-case $\widetilde O(n^{13/12+o(1)})$ private cut-release error.

LLM-agent defenses are typically evaluated one session at a time. In deployment, however, attacks can be distributed across independent agents, teams, and runtimes, leaving each local guardrail with only a sparse fragment. We formalize cross-agent asynchronous campaign attribution: linking sessions from the same latent adversarial campaign without shared runtime state, test-time campaign labels, or attacker identity oracles. We introduce Asynchronous Attribution Fingerprint Vectors ($A^2FV$), a lightweight proxy-side reference protocol for scoring pairwise campaign similarity from proxy-observable tool-use, timing, and prompt residue. We also construct SCD-v1, a controlled persona-matched benchmark with benign traffic, isolated attacks, multi-session campaigns, matched non-oracle evasion, and leakage audits. On SCD-v1, $A^2FV$ achieves 0.82 pairwise AUC for campaign linking, while score-only adaptations of per-session detectors and chunked LLM judges remain near chance under the same task. The strongest fixed signal is carried by structural and stylometric residue, while timing is retained as a diagnostic channel for richer proxy traces. Crossed-style controls show that the signal is partly style-sensitive but not reducible to style alone. Static and dimension-aware non-oracle stress tests further show that pairwise separability persists under controlled evasion. These results establish cross-agent campaign attribution as a distinct evaluation layer for securing LLM agents in the wild.

Fine-tuned foundation-model detectors dominate face-forgery benchmarks, yet they stay blind to generator families absent from training. We present GLID, a detector that repairs this blind spot with geometry instead of data. GLID treats the patch tokens of a single image as a sample from a manifold and estimates their local intrinsic dimension (LID) at several depths of a frozen vision transformer. This 12-dimensional, training-free signal enters a fine-tuned detector through a confidence gate whose strength is calibrated purely in-distribution. On a 16-axis cross-generator benchmark, GLID reaches 0.805 mean AUC, first among retrained state-of-the-art baselines and never significantly behind the strongest of them on any axis. It lifts the generation axes by +0.084 AUC while moving reenactment by only -0.005. Two empirical laws explain the design. First, forged faces bend the token manifold at family-specific depths: GAN artifacts peak at the last layer, diffusion artifacts peak mid-network, and the pattern survives four backbones, three dimension estimators, and non-face imagery. Second, fine-tuning absorbs auxiliary gains exactly where training data covers: injecting 1% target-family images erases a +0.100 gain, so geometric signals matter precisely where data is unavailable. The deterministic signal also cuts the cross-seed spread of accuracy 5.5x. Code, preregistered analysis gates, and per-image scores accompany the paper.

Large Language Models (LLMs) are increasingly fine-tuned for critical-domain Question-Answering (QA), yet choosing which small model to adapt, before paying the cost of adaptation, remains difficult. Fine-tuning can improve domain alignment, but it may also erode prior knowledge, weaken instruction-following, or increase hallucination, especially when labeled data are scarce or rapidly evolving as in cybersecurity. We present FiT (Find before Fine-Tune), a task-oriented diagnostic framework that characterizes small LLMs along three capabilities required for cybersecurity QA: vocabulary recognition, parametric knowledge, and contextualization of retrieved information. Using FiT, we conduct an empirical study of five open-weight 7-billion-parameter models under two fine-tuning regimes. We find that fine-tuning does not uniformly help: it consistently degrades vocabulary and parametric knowledge in small models, and the two regimes trade off differently. Knowledge-focused tuning causes moderate, rank-preserving degradation, whereas instruction-focused tuning collapses measured knowledge through induced abstention, inverting the knowledge ranking while leaving retrieval-grounded contextualization essentially intact. We quantify these regime-specific patterns with rank-correlation analysis and show that pre-fine-tuning FiT scores anticipate the direction of post-tuning change. Our results suggest that task-oriented diagnosis can screen out unsuitable models, avoid unnecessary fine-tuning, and support safer deployment of small LLMs in cybersecurity QA pipelines.

Public CVE descriptions report the technical conditions and impact of vulnerabilities, whereas the Auto-ISAC Automotive Threat Matrix (ATM) expresses an adversary's tactics and techniques. Because the two representations are not directly aligned, incorrect automated mappings in safety-critical environments may distort threat interpretation and mitigation prioritization, motivating a confidence-aware approach that distinguishes auto-confirmable mappings from uncertain cases. This paper reformulates automotive CVE-to-ATM mapping as a selective automation problem. The proposed framework generates candidate mappings via hierarchical in-context learning, then fuses self-consistency and LLM-based evidence verification signals into a calibrated meta-model. The resulting calibrated confidence score routes each candidate into AUTO, REVIEW, or HOLD. On the evaluation set, the proposed system substantially improved candidate-set precision at matched recall over a Flat zero-shot GPT-5.2 baseline. In the High-Confidence operating mode, the AUTO tier achieved a precision of 0.878, more than double the candidate-set base rate, and the calibrated confidence score achieved an AUROC of 0.868 in distinguishing correct from incorrect candidates. These results show that the framework can support selective automation by isolating auto-confirmable mappings from those requiring analyst review.

LLM-based browser agents are rapidly changing the threat landscape for web security. Unlike traditional automation frameworks that execute predefined scripts, these agents can autonomously navigate websites, reason about page content, and interact with web interfaces using natural-language instructions. This evolution raises fundamental questions about the effectiveness of bot management systems, widely deployed to defend against automated web abuse. In this paper, we present a systematic measurement study evaluating the resilience of both interactive challenge-based defenses and non-interactive trust-based defenses against two attacker classes: commercial Captcha-solving services and LLM-based browser agents. Our evaluation spans seven solver services and six agents, including cloud-hosted, self-hosted, AI-assisted, and browser-extension configurations, tested against hCaptcha, reCaptcha v2, reCaptcha v3, and Cloudflare Turnstile. Our results show that challenge-based defenses are broadly ineffective against commercial solvers, which achieve near-perfect bypass at negligible cost. The challenges can similarly be defeated by LLM-based agents when a dedicated solver module is available. Non-interactive defenses such as reCaptcha v3 exhibit stronger resistance, but our analysis reveals that this resilience does not reflect a fundamental security property. Through fine-grained interaction trace analysis, we find that two agents with nearly indistinguishable behavioral footprints yield divergent outcomes, one bypassing the defense and one failing, isolating execution-environment authenticity, rather than agent behavior, as the determining factor. These findings suggest that the security boundary of non-interactive defenses lies at the environment layer, with significant implications for how bot management systems are designed and evaluated.

Textual Collaborative Prompt Optimization (TCPO) extends TextGrad (Yuksekgonul et al., 2025) to a decentralized setting by allowing multiple clients to jointly improve prompts for large language models (LLMs) while keeping their data locally. Its reliance on free-form textual updating and aggregation introduces a new and largely unexplored attack surface, i.e., malicious instructions can be injected into local prompts and propagated through server-side prompt aggregation. Unlike conventional prompt injection attacks, attacking TCPO targets the collaborative optimization loop in TCPO. This setting is more challenging because malicious instructions must survive aggregation, persist through subsequent benign prompt optimization, and evade server-side defenses. To expose this risk, we propose Collaborative Prompt Injection (CPInj) attack that contaminates the aggregated global prompt with malicious instructions, degrades downstream task performance, resists purification by prompt optimization on benign clients, and evades advanced detection-based defenses on the server. We find that current defense methods are ineffective against CPInj. We further propose Anchored Purification Aggregation (APAgg), a defense-oriented aggregation that purifies malicious instructions without severely degrading TCPO utility. We conduct extensive experiments across three LLM families and five reasoning tasks in math, logic, and medicine, and demonstrate that our proposed attack reveals a critical vulnerability in TCPO. Although we take a first step toward mitigation, the attack remains highly effective and far from fully resolved, calling for more robust defense for TCPO.

Existing side-channel theories treat leakage as a holistic quantity $I(X;Y)$, without characterizing its internal structure. This paper studies the structural decomposability of encrypted-traffic side-channel leakage. Via the structural causal model $X\!\to\!Y_{\mathrm{size}}\!\to\!Y_{\mathrm{dir}}\!\to\!Y_{\mathrm{time}}$ and the mutual-information chain rule, total leakage is decomposed into three sequential increments for packet size, direction, and timing. Defenses are formalized as mechanism replacement by a strategy variable~$D$; coupling information $C_{\mathrm{size,dir}}=I(Y_{\mathrm{size}};Y_{\mathrm{dir}}\!\mid\!X)$ measures inter-dimensional dependence, and the Markov residual gives a testable condition for a single-dimension defense to sever downstream leakage. Causal efficacy~$\eta_d$ quantifies per-unit-cost suppression, and a Fisher-geometric approximation $I(X;Y)\approx\frac{1}{2\ln 2}\mathrm{Tr}(G\Sigma_\theta)$ holds under small perturbations. On the Wang dataset (95 websites), $Y_{\mathrm{dir}}$ dominates undefended leakage (0.637\,bits), while Tor's fixed 512-byte cells make $Y_{\mathrm{size}}$ degenerate; FRONT suppresses the direction term by 63%, yet its Markov residual of 0.021\,bits (95% CI $[0.016,0.027]$) shows it cannot sever timing leakage; $\eta_{\mathrm{dir}}=0.97$ vs. $\eta_{\mathrm{time}}\approx 0$ confirms FRONT's design intent. This yields a computable, structured leakage-accounting method for multi-dimensional joint defense design.

Deepfakes, especially face-swapping attacks, pose significant challenges to authenticity, security, and ethics across science, engineering, and society. While most existing detection/tracing approaches operate post hoc, proactive defenses that aim to intervene before deepfake generation remain limited in terms of real-world effectiveness. In this paper, we present PhantomSeal, the first proactive defense to simultaneously protect both the identity and the context of users' images from being used in face-swapping attacks, while supporting forensic tracing. We present a novel cloaking technique that embeds a selected identity as a stealthy identifier. This mechanism steers the deepfake generation process toward producing content that resembles the chosen cloak identity, thereby preventing successful face-swapping while enabling effective feature-based forensic analysis. The effectiveness and robustness of PhantomSeal is demonstrated in extensive experiments across different face-swapping architectures and models. For example, it reduces the attack success rate of SimSwap, an advanced deepfake model, to 0.30%, and correctly identifies 97.97% of manipulated content. The source codes is available at https://github.com/LiangqinRen/PhantomSeal.

The transition from stateless generative models in artificial intelligence to stateful, autonomous agents represents an architectural evolution that, while providing the capabilities of long-term planning and the automation of enterprise workflows, also represents the introduction of a new form of security threat, the Chronos Vulnerability. The Chronos Vulnerability represents the threat of memory-based attacks, including the Memory Injection Attack (MINJA) and the sleeper agent, in which the internal belief system of the autonomous agent is compromised, effectively decoupling the attack vector from the final catastrophic event. This study formalizes the threat model for persistence-based attacks and the threat of Dynamics Blindness in the context of the World of Workflows benchmark, demonstrating that traditional endpoint content filters are insufficient for the current stateful architecture. Consequently, this study synthesizes a defense-in-depth landscape, categorizing emerging frameworks such as diagnostic trajectory guardrails (AgentDoG), formal temporal verification (Agent-C), immunological memory consensus (A-MemGuard), and hardware-anchored trust via GPU-based Trusted Execution Environments (TEEs) and Zero-Trust memory architectures.