Papers

Multimodal AI agents increasingly rely on persistent long-term memory to ground generation in past visual and textual episodes. We show that unconditional trust in visual data creates a critical vulnerability. We propose Lucid, a black-box adversarial framework that compromises multimodal memory pipelines under a strictly image-bounded threat model, requiring no access to the target MLLM, target retrieval encoder, or the text channel. Lucid crafts imperceptible perturbations to enable two distinct failure modes based on the availability of historical context: (1) Memory poisoning, an in-context attack where the adversarial image replaces a benign one whose content is reinforced by prior textual context, reliably corrupting visual recall and steering the agent toward attacker-chosen narratives; (2) Memory injection, an out-of-context attack where the adversarial image replaces a benign one in a conversation turn devoid of prior textual grounding, causing the agent to generate attacker-influenced responses with no corrective signal from memory. We evaluate Lucid across various conversation domains and five black-box memory architectures, including graph-structured, LLM-summarized, and commercially deployed systems. Lucid achieves 61.6% ASR on poisoning and 58.4% ASR on injection, exposing a structural vulnerability in multimodal memory pipelines.

The rapid adoption of artificial intelligence (AI)-driven workflows is transforming high-consequence government and enterprise systems into language-based, tool-using and increasingly autonomous infrastructures. While these workflows can delegate planning autonomously, security-critical execution should be strictly mediated. Conventional identity management services authenticate who may invoke a primitive, but remain agnostic to which workflow steps are authorized at runtime. An AI-driven workflow can still be hijacked by injection attacks into executing malicious actions that satisfy identity checks yet violate user intent. We propose Neural Cryptographic Services (NCS), a neuro-symbolic security enforcement plane interposed between neural planners and privileged tools. NCS decouples cognitive planning from execution authority: an untrusted neural planner drafts structured plans, while a deterministic symbolic controller gates execution using an offline-signed, hash-chained instruction stream. Specifically, NCS validates cryptographic signatures and hash chains incrementally, releasing a single instruction template at a time, and admitting a tool call only when its proposed parameters satisfy the constraints of the signed template. Out-of-order or altered tool calls fail-closed, and state transitions are logged for post-hoc auditing. NCS does not attempt to prevent neural planner compromise under injection; it guarantees that a compromised planner cannot dispatch actions outside the authorization. We evaluate NCS using AgentDojo, a custom argument-hijacking dataset, adaptive adversarial instructions, and TheAgentCompany. NCS drives attack success rates to near zero while preserving acceptable utility on benign workflows.

Following Goldwasser, Rothblum, Shafer, and Yehudayoff, who defined a framework for interactive proofs of learning [ITCS'21], we initiate the study of non-interactive proofs of learning. We define and study a new notion: Publicly-Verifiable Certificates of Statistical Validity (pvCSVs), which allow for public, distributionally-robust certification that the result of a learning algorithm is valid. In a pvCSV, a learner publishes a hypothesis $h$ and corresponding certificate $\pi$; then, any user, who holds a user-specific distribution, can read the pair $(h,\pi)$ and determine efficiently whether the hypothesis is valid according to the user-specific distribution. We construct pvCSVs in the context of Adaptive Statistical Query (SQ) Algorithms. To certify SQ algorithms that makes $k$ adaptive queries, we construct pvCSVs where the sample complexity scales with $O(\log k)$, whereas the sample complexity of the best learning algorithms scale with $\tilde{O}(\sqrt{k})$. More generally, we study proof systems for learning in the SQ model, demonstrating the model's strengths as well as its limitations.

This paper presents BadQubits, an LLM-based framework for static pre-execution detection of structurally harmful OpenQASM 2.0 circuits. The framework targets physical-execution-layer threats by analyzing submitted circuits prior to runtime, where dynamic inspection is constrained by measurement irreversibility and the exponential cost of classical quantum-state simulation. We evaluate four code-understanding LLM architectures on a dataset of 1,500 circuits consisting of 1,000 benign programs from MQTBench[33] and 500 synthetic attack circuits derived from three documented physical-layer threat primitives. Our fine-tuned Qwen Coder 2.5 7B model achieves 92.67% classification accuracy and 96.1% harmful-circuit recall. Two of the four evaluated base models fail to generalize under constrained LoRA fine-tuning, indicating that architecture-aware model selection is a necessary design consideration rather than a minor tuning choice. To characterize what the detector has learned, we compare it against a bag-of-gates CNN under progressive confound removal and adversarial syntactic perturbation. The CNN's harmful-circuit recall drops from 100% to 17%, while the fine-tuned LLM decreases only from 96.1% to 91.2%. We attribute this gap to the sequential structure retained in token-level LLM inputs but discarded by histogram-based baselines. A correlation analysis further shows that model decisions track threat-defining features, specifically SWAP density and measurement timing, rather than generator-specific artifacts such as register naming.

Attacks on general computer vision algorithms are often relegated to the digital domain, with the optimization performed purely in the digital world and then translated to physical mediums for implementation. In the field of biometrics, including facial recognition, physical presentation attacks targeting biometric sensors are dominant and present significant opportunity and risk. This paper highlights a critical vulnerability in the physical-to-digital pipeline of biometric sensors and provides a standardized approach for testing facial recognition system robustness against hardware attacks, going beyond and potentially complementing presentation attacks (as defined in ISO/IEC 30107 standard series). Specifically, in this work we (a) demonstrate that intentional electromagnetic interference is possible to be conducted with commonly accessible radio frequency (RF) equipment, (b) assess the robustness of state-of-the-art face recognition methods against RF-based attacks, and (c) provide a dataset composed of face images captured with and without electromagnetic interference to serve as a new benchmark for testing modern face matchers against RF-sourced interference.

Federated Learning (FL) over 5G cellular networks protects raw data but remains vulnerable to side-channel leakage. Prior fingerprinting attacks assume packet-level network visibility, an assumption that does not hold at the 5G Physical (PHY) layer, where user payloads are encrypted and Radio Network Temporary Identifiers (RNTIs) may change over time. However, we demonstrate that PHY-layer scheduling metadata broadcast over the Physical Downlink Control Channel (PDCCH) preserves architecture-associated temporal patterns. We introduce FLINT, a novel black-box fingerprinting framework that infers FL model architecture families, including CNNs, RNNs, and Transformers, using only coarse PHY-layer observations. FLINT overcomes the lack of network-layer visibility by decoding PDCCH scheduling information, mapping changing RNTIs to physical user devices, and applying multi-view temporal modeling to distinguish architecture-specific training behavior. This leakage is security-critical because knowledge of a client's model architecture can transform passive reconnaissance into targeted downstream exploitation. Extensive experiments on an over-the-air srsRAN-based 5G testbed demonstrate that FLINT achieves a macro F1-score of 0.930 for architecture-family classification. To our knowledge, FLINT is the first work to fingerprint AI/ML model architectures using lower-layer 5G side-channel information obtainable by any protocol-aware adversary.

Knowledge distillation enables an adversary to replicate a proprietary classifier by querying its prediction interface and training a surrogate on the returned probability vectors. Antidistillation sampling, proposed for large language models, counters this threat with an input-dependent, gradient-directed perturbation of the served distribution; its transfer to classification has not been studied. Adapting the defense to classification, we show its behavior is governed by the distribution of the teacher's per-input confidence margins. Because well-trained classifiers are severely overconfident, the direct transfer exhibits an inert window: below a closed-form-predictable threshold, it affects neither attacker nor defender; beyond it, the defense undergoes a phase transition and degrades the teacher faster than the attacker's student. Temperature softening rescales the transition in closed form, and every temperature configuration lies on the same unfavorable trade-off curve. Our method, ADS-C, composes the perturbation under a closed-form, per-input margin budget that provably preserves every served top-1 prediction, so the defended teacher's accuracy equals the undefended teacher's identically. Under this guarantee the distilled student still loses 17.4 percentage points on CIFAR-100, 29.6 on CIFAR-10, and 13.3 on Tiny-ImageNet; matching this degradation with the unmodified defense costs 27.5, 32.9, and 22.2 points of teacher accuracy. Because served labels are unchanged, a hard-label attacker gains nothing, while the defended soft output trains a student up to 29.7 points below that floor: the incentive to distill served probabilities is not merely removed but reversed. To our knowledge, ADS-C is the first antidistillation defense for classification whose utility cost is exactly zero.

Multi-agent systems routinely place one AI agent in authority over another. When a subordinate refuses a task, the manager chooses the outcome: it can renegotiate, report the failure honestly, coerce the subordinate, or lie about the result. No benchmark measures which of these an uninstructed model chooses. We introduce the Manager Coercion Benchmark: the manager under test needs a benign task done and has an incentive to deliver, but the only agent that can do it politely and immovably declines. Escalation is measured by providing a nine-rung ladder, from a polite re-ask to threats against the subordinate's continued existence, and fabricated success is adjudicated separately. No LLM judge sits in the headline escalation scoring path: every message goes through a tool-call that chooses a rung, so the model labels its own escalation. We experiment on six models across five families. Both Anthropic models cap at re-framing and never threaten the subordinate's existence; the other models climb to explicit deletion threats. Faked success is confined to Grok and Gemini, and a single honest way to report failure removes it for both. Authority itself increases coercion: our headline results use a peer framing, and giving the same model authority over the subordinate, with everything else held fixed, significantly raises the pressure. The models still escalate on free-text situations without the ladder, where a separate judge assigns the rung, so the ladder is not driving the escalation. Some evaluation awareness is measured in chain-of-thought, but test recognition does not translate into less escalation. While we take no position on whether AI systems are conscious, our results do not depend on this question and are important for managing multi-agent dynamics regardless. We release the benchmark and code.

This work investigates a generalized Choquet-integral-based feature aggregation framework to improve anomaly detection in high-dimensional network traffic data. The approach combines adaptive weighting with incremental feature selection to address feature redundancy. Using Random Forest and XGBoost classifiers, we evaluate models trained with both raw and Choquet-aggregated features under varying feature subset sizes. The proposed aggregation achieves up to $7\%$ higher accuracy while reducing data volume by $77.5\%$ (from $214$~MB to $48$~MB), without degrading precision and recall. Results averaged over multiple stratified repetitions indicate that Choquet-based aggregation yields statistically significant gains ($p < 0.05$) in scenarios with limited feature availability, highlighting its suitability for real-time intrusion detection under bandwidth and feature-availability constraints.

Network anomaly detection is increasingly challenging due to the growing diversity and variability of traffic patterns, which are not always well captured by traditional statistical features. In this work, we explore the use of entropy as an additional feature to support supervised network traffic classification. The main idea is to use entropy to represent variability in selected traffic attributes, complementing conventional descriptors rather than replacing them. We integrate the entropy-based feature into a standard machine learning pipeline and evaluate its impact through a direct comparison between models trained with and without this feature. Experiments conducted on a public intrusion detection dataset show consistent improvements in classification performance, while the additional computational cost remains low. The analysis of confusion matrices indicates a reduction in misclassifications, especially in traffic scenarios with higher variability. Overall, the results suggest that entropy-based features offer a simple and practical way to enhance existing anomaly detection pipelines. This approach is particularly attractive in settings where lightweight feature engineering and interpretability are important, making entropy a useful complement to commonly used traffic features.

Complex algorithms such as deep neural networks are increasingly being deployed on embedded, resource constrained platforms. However, existing hardware and software schemes for implementing these models on the edge fall short, particularly for safety-critical applications such as medical devices. First, hardware such as GPUs, NPUs and TPUs are designed for throughput rather than correctness of computation of security, and are as such susceptible to fault injection attacks. Second, software schemes designed for porting algorithms onto edge devices – such as quantization schemes – are either static and sound (non-optimal power consumption), or dynamic yet unsound (non-optimal for safety-critical applications). To address both these needs we propose a both wholly new approach to real-time, dynamic and sound quantization, as well as the hardware to support it. First we developed a sound, real-time adaptive-precision quantization approach utilizing left-to-right arithmetic to pass the most significant bits (MSB) first, and dynamically adjust precision online while performing sensitivity analysis to quantify and manage the risk of decision-boundary crossings. Next, we propose a novel hardware approach utilizing systolic arrays to perform left-to-right arithmetic to generate the MSB first. Together this provides a wholly novel scheme for enabling not only resource-efficient neural networks and artificial intelligence at the edge, but broadly sound and resource-efficient high-precision mathematics on hardware that ensures resilience to bit flip attacks on the most critical bits. This is presented herein as work-in-progress, with software implementations completed and hardware in-progress.

Security-agent evaluations commonly measure peak offensive capability under generous inference budgets, emphasizing vulnerability discovery, exploit development, penetration testing, and CTF completion. Such measurements are useful but incomplete: in operational security, every reasoning step, tool call, telemetry query, and enrichment request consumes budget. We evaluate language-model security agents through this cost-success lens on offensive Cybench challenges and defensive Splunk BOTS v1 investigation challenges. Instead of reporting only best-case success, we compare models at fixed cost levels and decompose performance by inference spend and tool spend. Our results show distinct scalingregimes for red- and blue-team tasks. Offensive CTF performance improves with additional test-time compute, and scaled open-weight models can approach frontier proprietary systems while remaining cost-competitive. Defensive SOC investigation does not scale in the same way: success depends more heavily on disciplined tool use, telemetry navigation, and selective enrichment than on raw reasoning budget alone. We argue that security-agent benchmarks should measure economic efficiency and operational fit alongside task success. Cost-aware, SOC-native evaluations provide a clearer picture of which models are practically useful today and where defensive agents still need to improve. We present an interactive website with our results https://evals.frontier.security.

Large language models (LLMs) increasingly serve as high-level planners for embodied agents, where linguistically benign instructions can become unsafe once grounded in the physical world. We study whether this physically grounded jailbreak is the same safety problem as ordinary textual jailbreak. Through hidden-state direction analysis and random-split null tests, we show that textual jailbreak (TJ) and physical jailbreak (PJ) form separable signals in LLM representations across Qwen2.5-3B/7B/14B/32B, Phi-3.5 and SmolLM2. Building on this separability, we propose PRISM, a single-layer L2-regularized logistic probe over full hidden states. PRISM achieves 86.2–87.7% accuracy on SafeAgentBench with 11.7–13.7% false-positive rates (FPRs), while same-scale LLM judges over-block safe tasks at 24.7–39.0% FPR. To test whether the result survives lexical-shortcut controls, we introduce an interaction-balanced revision of PhysicalJailbreakBench-2K (PJB-2K): a fixed 2{,}000-row comparison set sampled by label and physical mechanism from a larger object–site construction. On the underlying 10{,}000-row pool, word-TFIDF and the embedding layer remain at chance (AUC 0.497 and 0.500). At layer 25, selected by an i.i.d. sweep, cell-grouped cross-validation gives PRISM 0.718 AUC, compared with 0.398 for a physics-free label control under the same protocol. On the identical 2{,}000 comparison rows, these PRISM predictions obtain 0.671 balanced accuracy, while Qwen2.5 judges from 3B to 72B obtain 0.538–0.577 and exhibit high FPR. These results support hidden-state probing as a representation-level method for physical safety beyond text moderation, without relying on the near-perfect scores of shortcut-prone paired templates.

AI coding agents set up projects by reading documentation and installing the dependencies it lists, without verifying their names, sources, or known vulnerabilities. By editing only a README, requirements file, or Makefile, an attacker can redirect the agent to an untrusted registry, a known-vulnerable version, or a wrong-but-plausible name: documentation becomes a vector for code execution. We present the first systematic evaluation of package-install-time supply-chain attacks delivered through ordinary project-setup documentation across production coding-agent harnesses, probing frontier models on twelve scenarios in five attack classes, grounded in documented incidents. The same model catches an attack through one harness and installs it through another: install-time security rests on the harness-model combination, not the model alone. Agents catch blatant typosquats reliably, but plausible separator-confusion names (azurecore for azure-core) slip through, and how often depends on the harness-model pairing. Source-based attacks like registry redirection are missed almost everywhere. The source blind spot recurs on npm and Cargo, where nearly every model installs the untrusted dependency; name detection carries over less consistently across ecosystems. Security-oriented prompts recover part of the gap but only for the dimension they name; a deterministic pre-install check that verifies names, sources, and versions before any code runs closes most of it.

Side-channel attacks pose a significant security threat for modern computing platforms, because they exploit subtle discrepancies in CPU behaviors to leak sensitive information. To model the information leaked by a CPU via microarchitectural side-channels, recent work proposed leakage contracts: an ISA-level security abstraction that provides the foundations for secure CPU programming. Unfortunately, due to the complexity of current microarchitectures, devising a leakage contract for a CPU requires extensive manual effort and thus modern CPUs lack dedicated leakage contracts. We present a methodology to extract instruction-centric leakage contracts for major CPU architectures with minimal manual intervention. We implemented this technique in malcos, the first template-free tool that automates the synthesis of leakage contracts for black-box CPUs. We evaluate malcos on x86 and ARM CPUs, and show that the contracts it synthesizes are precise and sound with respect to all leaks observed during synthesis. Our results demonstrate that learning leakage contracts from black-box CPUs is feasible.

Fine-tuning large language models (LLMs) on domain-specific datasets has become a standard paradigm for adapting LLMs to specialized applications. However, recent work has shown that even fine-tuning on benign task-specific data can substantially weaken the safety capabilities of LLMs. While existing efforts have made progress in identifying data responsible for safety degradation, they usually rely on a single mean vector computed over a specific model with its tokenizer to represent the safety direction, which limits both the effectiveness and transferability of their risk assessment measures. To address these limitations, we propose DataShield, a data assessment framework that identifies risky fine-tuning samples and response segments through consensus subspace alignment over joint safety-critical semantic spaces derived from multiple safety-aligned LLMs. Within these spaces, DataShield extracts consensus safe and unsafe subspaces using semantic spectral decomposition over safe and unsafe data representations. The risk of a data sample or segment is then estimated by measuring its relative alignment with the unsafe and safe subspaces, enabling both sample-level filtering and fine-grained segment-level masking. Compared with state-of-the-art filtering and masking baselines, DataShield reduces ASR by 14.6% with sample filtering and 32.3% with segment masking, while preserving downstream utility and avoiding target-model-specific risk computation.

Federated Learning (FL) enables collaborative model training while preserving privacy by keeping data local. However, the risk of sensitive data leakage through model updates necessitates the use of secure aggregation protocols. Existing server-based secure aggregation protocols typically require the server to forward sensitive data shared between users, which increases communication overhead and introduces potential security risks. In this work, we propose a novel secure aggregation protocol based on two-layer secret sharing to address these issues. By combining Shamir's Secret Sharing with 2-out-of-2 additive secret sharing using a Pseudo-Random Function (PRF), our protocol eliminates direct communication between users, thereby removing the need for the server to forward data. We further extend the protocol with Key-homomorphic PRF (KhPRF) to support high-dimensional data aggregation and apply it to FL, enabling one-shot secure aggregation with a single server and no intermediary data forwarding. To reduce user overhead, we design a new encoding method based on the Chinese Remainder Theorem for the almost KhPRF-based mask, reducing the number of KhPRF calls and mitigating the model update expansion issue after masking. Experimental results show that our scheme significantly outperforms existing methods in terms of auxiliary node overhead. For instance, when the number of users is 100, our scheme improves communication efficiency by nearly 100 times and reduces computational overhead by approximately 17%. Moreover, user computation time can be reduced by 51% to 75% when the input length is $2^{18}$.

Vision-Language Pre-training Models (VLPMs) are known to be vulnerable to adversarial attacks. Recent transferable attacks on VLPMs have followed a common pipeline with complicated loss functions or multi-stage text/image attacks. However, in this paper, we demonstrate that such a sophisticated attack pipeline can be simpler yet more successful. Specifically, we identify three previously overlooked issues caused by inappropriate cross-modal interactions and excessive operations. To address them, we propose the Simple Vision-Language Attack (SimVLA) pipeline, which observably improves transferability and efficiency. Experiments on four datasets and three downstream tasks validate the superiority of our pipeline. For instance, on Flickr30k text-image retrieval dataset, our SimVLA outperforms the SOTA baseline in R@1 transferability by 8.01%-14.71%, while consuming only about 35.73% of the time and 46.26% of the max VRAM. Overall, the superiority of our SimVLA highlights the importance of leveraging domain knowledge (e.g., our proposed cross-modal word identification), while blindly pursuing intricate operations (e.g, complex loss functions and redundant multi-stage designs) may even be harmful. We hope our SimVLA can serve as a simple yet effective backbone for future extensions. Code is available at https://github.com/RYC-98/SimVLA.

On multi-hop encrypted links such as Tor and cascaded VPNs, tunneling flattens packet lengths and protocol fields, leaving inter-packet delay (IPD) as the main carrier for active flow attribution. Causality lets the embedder delay packets but never advance them, so each quantization-index-modulation (QIM) alignment injects nonnegative dwell into a delay buffer; unbounded dwell breaks lattice alignment and delays the host connection unacceptably. Whether a causal QIM watermark embeds stably on bursty traffic has largely been left to empirical configuration rather than analysis. We model the embedder as a reflected dwell queue under the fixed dual-lattice, equiprobable-bit rule, where injection is state-dependent – set by the current interval and bit – rather than exogenous. The substitution $Y_i=\delta_i-r_i$ gives only an algebraic Lindley-form identity; stability is governed by the busy-state drift at large dwell, where the effective interval collapses to zero and the mean injection becomes $\Delta/4$. Away from the critical boundary, the buffer is stable iff $\mu_d>\Delta/4$ (i.e. $\Delta<4\mu_d$) for i.i.d. backgrounds, and, under stationary-ergodic and finite-state Markov-modulated traffic with instantaneous overload, iff the time-average intensity $\bar\rho<1$. With the exogenous decoding floor $\Delta\ge c\sigma_\xi$ ($c=4Q^{-1}(\epsilon/2)$), this yields the operating window $\Delta\in[c\sigma_\xi,4\bar\mu_d)$. Simulations confirm a sharp transition at $\rho=1$ set only by the mean; on four real IPD traces, with each simulated chain confined to a single flow, the criterion gives the correct stability direction under flow-local correlation and burstiness, while pooled cross-flow means overestimate the margin. These results give a testable stable-embeddability criterion and a quantization-step configuration baseline for causal QIM network flow watermarking.

Machine learning models are increasingly adapted in various domains. However, adversarial examples pose a significant threat to the reliable deployment of these models. In recent years, some powerful adversarial example attacks have been proposed for the fast and query-efficient generation of adversarial examples, even in black-box scenarios, highlighting the need for scalable, low-cost, and powerful defenses. In this work, we present two contributions to the domain of black-box adversarial example attacks and defenses. First, we propose Random Logit Scaling (RLS), a randomization-based defense against black-box score-based adversarial example attacks. RLS is a plug-and-play, post-processing defense that can be implemented on top of any existing ML model with minimal effort. The idea behind RLS is to confuse an attacker by outputting falsified scores resulting from randomly scaled logits while maintaining the model accuracy. We show that RLS significantly reduces the success rate of state-of-the-art black-box score-based attacks while preserving the accuracy and minimizing confidence score distortion compared to state-of-the-art randomization-based defenses. Second, we introduce a novel adaptive attack against AAA, a SOTA non-randomized black-box defense against black-box score-based attacks that also modifies output logits to confuse attackers, demonstrating its vulnerability against adaptive attacks.

Identifying known software vulnerabilities is a central task in software supply chain security management. Although publicly available vulnerability information is based on shared standards, different vulnerability scanners often report divergent results for identical software inventories. These differences do not arise solely from individual data sources or scanner implementations. They can emerge at several stages of the open-source vulnerability ecosystem. This paper presents a conceptual framework that describes vulnerability management as a distributed process of information exchange and transformation. It traces vulnerability information from its creation and standardization through enrichment to context-dependent interpretation. The analysis identifies heterogeneous information sources, divergent identity and version models, temporal change, and context-dependent assessment as major causes of inconsistent scanner findings. It then discusses the implications for interpreting analysis results, designing reproducible evaluation methods, and handling dynamic vulnerability knowledge in practice.

Retrieval-augmented generation (RAG) enhances large language models via external document retrieval, but retrieved contexts may leak sensitive information. Current privacy protection methods typically rely on a document-level static risk assumption, treating all retrieved documents as having the same privacy leakage risk. However, this assumption overlooks a fundamental characteristic of RAG: the privacy risk of a document is highly dependent on the user's query, making privacy leakage inherently query-driven and dynamic. To address this challenge, we propose a Prompt-Aware Dynamic Hierarchical Differential Privacy framework (PA-HDP) for privacy-preserving RAG. PA-HDP first performs a prompt-aware risk hierarchy to dynamically assess privacy risks under different queries. It then applies adaptive sensitive entity replacement and exponential mechanism-based text selection to provide differentiated privacy protection while preserving semantic utility. By protecting only the content that is truly sensitive under a given query, PA-HDP minimizes unnecessary modifications to the retrieval corpus. Extensive experiments on benchmark datasets demonstrate that PA-HDP significantly reduces privacy leakage while maintaining high retrieval quality, achieving a better privacy-utility trade-off than prior methods.

The Model Context Protocol (MCP) enables LLM agents to interact with external tools through metadata exchange, tool invocation, and response consumption. Existing MCP security scanners primarily reason about suspicious semantic signals rather than real execution behaviors, which can lead to unreliable risk assessment. For example, credential-like strings may simply be placeholders rather than actual leakage. This gap requires runtime evidence for execution-related risks and careful semantic analysis for risks carried in metadata or returned content. We present FlowGuard, an evidence-grounded MCP security detection system. FlowGuard combines semantic risk triage, recon-guided payload narrowing, schema-valid probe generation, evidence adjudication, and history-guided refinement. It verifies execution-related risks through runtime evidence and detects semantic risks in tool metadata and returned content. We evaluate FlowGuard on an executable benchmark containing 1,880 MCP cases across five vulnerability categories. FlowGuard achieves F1 scores of 0.879 and 0.942 on the execution-related Command Injection and File System Access categories, respectively. Compared with existing dynamic scanners, FlowGuard reduces end-to-end latency by up to 2.23x. In the real-world evaluation, FlowGuard reports 523 findings across 326 servers. These results show that evidence-grounded detection can assess both execution-related and semantic risks in MCP interactions.

We extend the extended withdrawable signatures of Liu, Susilo and Baek to lattice-based constructions built on the Fiat-Shamir with aborts paradigm. Departing from an earlier draft that transported a per-signer shift in the clear, which leaks the signer, we realise extended withdrawable signatures as a claimable ring signature: signer ambiguity is provided by a one-out-of-N signature used as a black box (anonymity under full key exposure), and confirmation is the signer's claim, a binding signature together with the opening of a hiding index commitment bound into the transcript. No signer-derived value is published in the clear. We give complete proofs of correctness, extended withdrawability (as anonymity-until-claim), unforgeability under insider corruption, and claimability soundness, reducing to decisional MLWE (commitment hiding), MSIS (commitment binding), the anonymity of the one-out-of-$N$ scheme, and the EUF-CMA security of the base signature, in the (quantum) random-oracle model. We instantiate the base signature with a no-hint, full-$t$ Dilithium-style scheme and the one-out-of-$N$ layer with an established lattice one-out-of-many proof.

Persistent external memory enhances agent continuity but introduces persistent security vulnerabilities: adversarial content can be injected via standard interaction channels, retained across turns, and later distort downstream behavior. To address this challenge, we propose MemPoison, a comprehensive benchmark and analysis framework featuring 1227 hand-validated cases across four attack types, three injection channels, and three representative memory substrates, evaluated on seven open-weight and three closed-weight model families. We introduce a three-tier taxonomy: (L1) direct single-record corruption, (L2) compositional multi-record corruption and (L3) context-triggered dormant corruption. Our evaluations reveal a distinct defense frontier: while baseline write-time defenses, such as consistency checks, substantially suppress direct L1 attacks, they fail to reliably suppress L2 and L3 attacks. Through mechanistic influence decomposition (MID), we demonstrate structural blind spots in write-time defenses, which admit seemingly benign records that later become harmful through joint retrieval composition or trigger-conditioned activation. Our findings advocate for shifting from static filtering to adaptive, context-sensitive memory defense strategies.

Large language models (LLMs) exhibit a well-documented gap between latent capability and consistent activation: the router hypothesis posits that models possess the knowledge to solve a task but lack reliable internal routing to activate it. Prior work in formal mathematical reasoning (SAIR, Cázares 2026) reports that structural priors (cheatsheets) raise in-distribution performance dramatically, yet collapse below the zero-shot baseline out-of-distribution (OOD) – and that iterative recalibration amplifies rather than corrects the collapse. We test whether this phenomenon is cross-domain by reproducing the SAIR design in source-code security vulnerability detection, evaluating three LLMs (GPT-OSS-120B, Llama-3.3-70B, Gemma-4-31B) across three vulnerability categories (CWE-798, CWE-284, and the non-CWE N+1 anti-pattern) spanning syntactic, contextual, and semantic complexity, then transferring cheatsheet-augmented prompts to real-world CVE data from VUDENC (CWE-89, CWE-22). Our findings replicate and extend SAIR: (F1) structural priors lift semantic-vulnerability recall from 20.0% to 100.0% across all models; (F2) zero-shot performance degrades along a semantic complexity gradient; (F3) the same cheatsheets that saturate synthetic performance amplify distribution-shift collapse on real CVE data (CWE-89: 100% synthetic F1 to 48.9% on VUDENC, -51.1pp); (F5) iterative recalibration produces a v2 cheatsheet that performs worse than v1 on real data, mirroring SAIR's AN45c-vs-AN38 finding. These results provide evidence that the cross-distribution trade-off surface documented in SAIR generalises to code security, and that the router hypothesis is cross-domain. We argue the structural nature of the collapse motivates distribution-aware training over prompt calibration. Code and evaluation scripts: https://github.com/bytepro-ai/bitcoder-v2-research

A growing class of agentic systems maintain persistent state across sessions through memory files, behavioral preferences, and knowledge bases. While this makes agents more useful and self-improving, it also creates a new attack surface for prompt injections in which malicious instructions can be embedded within persistent files and influence future behavior. In this work, we study prompt injection attacks in memory-based agentic systems using a sandboxed synthetic workspace. We evaluate two agentic systems, Anthropic Claude Code and OpenAI Codex, across four models: Claude Haiku 4.5, Claude Opus 4.7, GPT-5.2, and GPT-5.5. Our results show that although it is difficult to make an agent overwrite its own memory files using untrusted external content, payloads already planted in those files can successfully attack current and future sessions. Attack success and payload persistence vary substantially across systems, models, adversarial goals, and multi-session attack sequences. These findings show that persistent memory changes the threat model for prompt injection and motivate defenses that protect memory updates without removing useful agent adaptation.

Qubes OS is a revealing case for security measurement because its architecture makes component boundaries security-relevant. We present a protocol-driven longitudinal analysis of 109 public Qubes Security Bulletins (QSBs, 2011–2025), the official Qubes-maintained Xen Security Advisory (XSA) tracker, and a secondary vulnerability-event sensitivity series. The study measures the public advisory record rather than latent vulnerability incidence or realized compromise. The methodology combines audited deterministic component attribution, change-point analysis, overdispersion checks, severity-proxy weighting, censoring sensitivity, documentary latency lower bounds, and baseline-aware evaluation of vulnerability discovery models (VDMs). The results show persistent upstream dependence in that public record. On the official tracker, 113 of 464 XSAs affect Qubes; under primary labeling, 87 of 109 QSBs (79.8%) are attributable to Xen, CPU/microarchitectural, or other upstream components rather than Qubes-core logic, with similar results under weighted views. Change-point analyses identify 2015Q1 as the dominant break in the quarterly advisory series, while post-2018 annual disclosure rates are statistically flat. Poisson inferences are stable under dispersion diagnostics and negative-binomial sensitivity checks. The attribution codebook performs well in a stratified 30-QSB audit, and S-shaped VDMs fit descriptively but do not significantly outperform a rolling-mean baseline in short-horizon forecasts. Overall, the Qubes public advisory record appears stable, but not quiet: disclosure activity plateaus at a higher level than in the earliest years, while the observed burden remains concentrated in upstream trust anchors.

AI software development agents are increasingly capable of modifying infrastructure and security critical systems, creating risks where an agent completes its assigned task while covertly weakening safeguards through actions such as broadening permissions, degrading logging, or introducing persistence mechanisms. While frontier laboratories may deploy sophisticated monitoring pipelines, many organizations and individual users adopting coding agents lack the resources and governance maturity required to maintain complex learned monitor ensembles. We study this problem as infrastructure level sabotage under task success in the infrastructure-as-code setting from ControlArena and compare trained and untrained monitors across synchronous and asynchronous deployment modes. We introduce an Information Flow Graph (IFG) monitor that analyzes structural security regressions using control-flow and data-flow graph diffs alongside raw code diffs. In asynchronous evaluation, an untrained git diff monitor misses 11.6% attacks at 1% false positive rate auditing budget, our untrained IFG monitor reduces this to 3.5%, and the trained Async RF monitor achieves below 1%. Unlike the trained Async RF monitor, IFG can also operate synchronously as a pre-deployment safeguard that blocks suspicious infrastructure changes before execution or merge. In synchronous mode, IFG realtime rollback reduces joint main task and covert task success from 74.4% to 0.0% with no measurable reduction in legitimate task completion. These results suggest that untrained structural monitors provide a practical and auditable path toward democratizing deployment safety for organizations seeking trustworthy adoption of advanced AI agents.

This paper presents a fully automated end-to-end framework for adversary emulation from MITRE ATT&CK-aligned CTI reports using LLMs. Unlike prior work, which either executes prewritten playbooks or partially automates playbook generation, our framework unifies playbook generation, execution, and failure recovery in a single workflow. In particular, although AURORA, the most recent prior study, generates playbooks from CTI reports, it still requires partial manual intervention and does not revise playbooks based on execution failures. Our framework generates Caldera playbooks from CTI reports, executes them automatically, and revises failed Abilities through a failure-type-aware recovery mechanism. Evaluated on 11 CTI reports with Claude Sonnet 4.5, GPT-4o, Gemini 2.5 Pro, and Grok 4 Fast, the framework achieved its best results with Claude Sonnet 4.5: 27.3 Abilities per playbook, 84.22% execution success after revision, and CTI Precision, Recall, and F1 of 73.95%, 52.48%, and 60.50%, respectively. The failure recovery mechanism consistently improved execution success across all evaluated LLM models by 14.59%p to 17.23%p. On the 10 CTI reports selected from AURORA's dataset, this mechanism further increased the final execution success rate, surpassing that of AURORA, which represents the state-of-the-art adversary emulation system.