Papers

Leading large language model providers now conceal their models' step-by-step reasoning, or chain-of-thought, to protect intellectual property and limit information leakage. Rather than storing these traces server-side, providers return them to the client as blocks of encrypted text, which the client passes back with each subsequent request. Building on prior research, we identify an architectural vulnerability: these encrypted blocks are fully compatible and interchangeable across different sessions, users, and models within a provider's ecosystem. We exploit this compatibility to develop a scalable decryption jailbreak. By injecting an encrypted reasoning trace from a given model into a weaker, and less safeguarded model from the same provider, we force it to decode and output the trace verbatim in plaintext, without ever jailbreaking the more capable model directly. This vulnerability enables four distinct attack vectors. First, it circumvents anti-distillation mechanisms, allowing adversaries to extract a proprietary model's reasoning, as we demonstrate across Anthropic, OpenAI, and Google. Second, it allows for large-scale private data extraction. Developers frequently share session logs publicly, unaware of contents of the encrypted blocks. By decoding 315,320 reasoning blocks scraped from public repositories, we recovered 367 Personally Identifiable Information (PII) artifacts and 182 credentials. Third, it inadvertently reveals hazardous information hidden within the reasoning process, even in cases where the model's final, visible output safely rejects a malicious request. Fourth, attackers can leverage this flaw to execute invisible prompt injections, embedding malicious payloads entirely within encrypted blocks to poison public agentic rollouts. Following responsible disclosure, we propose concrete cryptographic and system-level mitigations to secure client-side reasoning.

Request for comments (RFCs) are Internet standards, memorandums, and related technical documents about core Internet protocols made via and released by the Internet Engineering Task Force (IETF). In the early 1990s each RFC was required to have a section for security considerations. The present work examines these sections. According to the empirical results, (1) over 90% of the RFCs sampled have discussed security explicitly in these sections, (2) although mandatory security requirements have only seldom-if ever-been imposed. Furthermore, (3) the RFC-to-RFC reference network specific to the security consideration sections is sparse, although a few RFCs and their security consideration sections are heavily referenced. In addition, (4) the volume of references peaked during a period from circa mid-1990s to mid-2010s. Regarding the topics discussed in the sections, (5) these do not represent general security issues, such as spoofing or eavesdropping; rather, the topics mostly reflect distinct security issues specific to distinct protocols. With the exceptions of network security in general, security specifications, and routing, (6) also the longitudinal evolution of the topics is protocol-specific. As the subject matter has not been previously examined, these empirical results fill a gap in the standardization literature.

Network traffic analysis faces significant challenges with encrypted communications, primarily due to limited visibility into packet contents and the inherent imbalance in available datasets, particularly for anomalous traffic patterns. This paper addresses these challenges by exploring Generative AI (GAI) techniques to create realistic and balanced synthetic encrypted traffic datasets. Our approach incorporates feature analysis, clustering-based data generation, and comprehensive classifier evaluation to ensure synthetic data quality. We demonstrate that properly generated synthetic data can effectively supplement real- world datasets, achieving up to 93% performance when training classifiers compared to those trained on real data. The proposed methodology preserves critical statistical properties and feature correlations while enabling the creation of balanced datasets, ad- dressing the persistent challenge of anomaly underrepresentation in cybersecurity data. Along with the results we provide complete programming code designed and implemented in this work.

Agent skills are emerging as an important attack surface in LLM-based agent systems. Through an empirical study of existing skill scanners, we find that current defenses mainly inspect individual skills, leaving risks from cross-skill composition insufficiently examined. This creates a practical blind spot: multiple locally plausible skills may pass security checks while collectively forming a harmful workflow during agent execution. To investigate this threat, we propose ColluSkill, a collusive multi-skill-chain attack framework that decomposes a complete malicious intent into interdependent sub-payloads embedded in independently packaged skills. The attack does not rely on any single malicious skill, but emerges from the ordered composition of locally plausible behaviors through contextual dependencies, artifact passing, and execution handoffs. ColluSkill further employs LLM-based chain planning and scanner-feedback refinement to preserve chain-level attack semantics while reducing suspicious signals in individual sub-skills. To defend against such attacks, we propose ChainGuard, a context-aware skill-chain scanner that jointly analyzes a candidate skill and the skills already installed in the agent environment. ChainGuard reconstructs cross-skill dependencies, artifact flows, capability compositions, and downstream behaviors to identify risks that emerge only at the workflow level. Experiments on six representative skill scanners show that ColluSkill achieves an average attack success rate of 96.0% and consistently outperforms the evaluated single-skill and multi-skill attack baselines. Meanwhile, ChainGuard reduces the attack success rate to 22.5% while allowing 99.5% of benign workflows to pass, highlighting the importance of chain-level security analysis for agent skill ecosystems.

We give a full-key-recovery attack on MQOM v2.1, a Round-3 candidate in the NIST additional-signature process, that recovers the complete signing key from one accepted signature and uses it to sign a fresh message. If $\delta=\operatorname{FirstBits}_{\lambda}(x)$ is the prefix of the witness $x$, the sibling path determines a public value $A$ such that tree parity gives $s=\delta\oplus A$. Substitution into the hidden-leaf commitment yields $$\mathsf{Enc}_K(\delta\oplus A)=T\oplus\mathsf{LinOrtho}(\delta)$$ with public values $K$ and $T$. The correction in the same signature expands a solution into a complete witness, while the public MQ relation identifies those yielding valid signing keys; serializing such a witness gives the secret key, enabling a fresh-message signature accepted by the reference verifier. We evaluate this equation over the specified AES/Rijndael circuits using retained circuit state along a Gray traversal. Complete-domain scans for Categories I and V cost $2^{142.335112}$ and $2^{271.794162}$ Boolean gates. Category-III scans cover $1/2+2^{-20}$ and $0.580004770183$ of the domain at costs of $2^{206.774558}$ and $2^{206.988685}$ gates. All four totals are below the NIST security benchmarks. Reduced-domain runs against the reference implementation recover the byte-exact witness and key in all three categories and produce a fresh-message forgery accepted by the reference verifier. Independently generated source-syntax circuits evaluate the fixed ciphers over the stated domains and translated L3 prefixes, while an exact ideal-cipher factorial-moment bound controls additional equation preimages passed to public-key validation. Every value in the equation is fixed by the accepted transcript, so salt-bound global-root expansion changes its public constants without removing the one-signature recovery channel.

AI coding agents now write a growing share of production code, and human security review does not scale at the rate code is generated. The agents in widest use are closed-weight, so a deploying team cannot read their internals. It can instead run an open-weight model as a reviewer over the agent's output. That reviewer's activations are readable. We ask whether reading those activations recovers a security signal that simply asking the same reviewer misses. We fit a single linear probe per model on a corpus of paired vulnerable-and-fixed Python functions, then test it without retraining on real disclosed vulnerabilities whose weakness type the probe never saw in training, across five open-weight reviewer models. On the vulnerabilities fixed by changing a single function, the probe scores the vulnerable function above its fix on 61-67% of cases for every model, beating the 50% chance line. It also beats the same model's prompted YES/NO win-rate read from its logits, under every prompt we try. Asking the model for a written verdict, even with chain-of-thought, returns the same answer on the vulnerable and fixed function most of the time and so cannot tell them apart. Model activations carry a code-security signal that prompting the same model misses.

Internal safety scores judge a prompt before any text is generated, and they are validated by how well they separate harmful prompts from benign ones. That separation is then read as evidence that the score will also catch the attacks that succeed. Harmful intent is a property of the prompt. Jailbreak success is an outcome produced later by a particular target model, decoding policy, and judge. A filter tuned on a score that measures the wrong quantity spends its false positive budget on attacks that would have failed anyway. In this paper we audit that inference. Attention based measurements are usually read from prompt dependent locations, so a wrapper changes both the content being judged and the place the signal is taken from. We therefore introduce Active Attention Probing, which supplies a fixed content independent measurement coordinate. We pair every base goal with a plain and a wrapped version and generate real completions from the target models. On Llama, wrapping raises harmful generation from 0.05 to 0.27 while harmful intent AUROC falls from 0.936 to 0.803, so the attacks grow more dangerous while the prompts look safer to the score. Among wrapped harmful prompts the outcome AUROC is 0.220, which places the attacks that succeeded below the attacks that failed. Rare token, passive, and detector derived channels reproduce the reversal on the same matched design, and the reversal itself persists across three target models, seven attack families, and two independent judges. Distribution shift then degrades calibration and threshold transfer before it degrades ranking.

Security research artifacts—repositories, PoC exploits, and validation pipelines—are increasingly produced by LLM/agent-driven vulnerability workflows, yet the gap between publicly available, runnable, signal-producing, and semantically confirmed artifacts is poorly measured. We conduct a pre-registered reproducibility audit of this literature. A search covering 2023–2026 with dual screening yields a 104-paper consensus corpus, of which 59 papers (56.7%) have a publicly reachable artifact. We execute an 18-paper sample at R0/R1 and all 102 cases of the anchor benchmark (arXiv:2509.24037), with patched-counterfactual verdicts on 30 signal-producing cases and matched-negative-control verdicts on 19. Three findings stand out. First, 58/102 (56.9%) anchor cases contain a script-internal CVE identifier that diverges from the declared directory CVE. Second, only 10/18 (55.6%) paper-level artifacts complete their declared workflow at R0, rising to 11/18 (61.1%) after environment-only R1 repair. Third, artifact-embedded oracles prove unreliable: 20/30 patched-counterfactual audits still produce the claimed signal on the patched build, 7/19 matched negative controls still trigger on benign input, and the oracle confusion matrix has sensitivity 60% and specificity 45%. A trigger on the vulnerable build is not evidence of CVE-specific reproduction without a clean patched counterfactual. These are exploratory results from a pre-registered protocol, and our protocol—pre-registered post-conditions, R0/R1 repair ladder, G1–G3 semantic evidence levels, and patched-counterfactual oracles—is a reusable template for the security reproducibility community.

Large reasoning models (LRMs) achieve remarkable success on complex tasks but remain vulnerable to harmful prompts that induce unsafe outputs. Recent methods align LRMs using direct refusals or safety rationales, yet often focus on prompt patterns rather than intrinsic attack mechanisms. As a result, these pattern-centric alignments struggle to generalize across diverse jailbreaks, compromising adversarial robustness and reasoning utility. We propose AdvSafe, a dual-adversarial framework that enables LRMs to internalize unsafety knowledge by explicitly deconstructing adversarial mechanisms. This moves beyond pattern-dependent traces, fostering robust cognitive defense without compromising reasoning utility. Our pipeline operates via a two-phase adversarial game. First, in adversarial synthesis, an autonomous agent dynamically crafts deceptive jailbreak prompts, adapting its strategies to breach a strong teacher model. Second, in adversarial extraction, the breached teacher executes a cognitive counter-attack. For every successful jailbreak, the teacher unmasks the camouflage, explaining why the attack succeeds and how such prompts can be identified and mitigated. This dual-adversarial process yields a compact reasoning dataset capturing rich, generalizable unsafety knowledge. Student models trained on this dataset implicitly acquire safety alignment through intrinsic threat comprehension. Experiments show that with only 1K synthesized samples, AdvSafe-aligned LRMs achieve significantly stronger jailbreak robustness than existing baselines, with almost no utility degradation. Furthermore, AdvSafe improves robustness against out-of-distribution prompts, demonstrating that learning unsafety knowledge enables a superior robustness-utility trade-off and generalizes beyond seen attack patterns.

Real-world cyberattacks often require sustained progress across multiple hosts and network segments, making multi-hop cyber ranges essential infrastructure for studying and improving LLM agents' ability to sustain complete attack chains. Prior work has scaled isolated vulnerability tasks and constructed multi-host scenarios from manually specified vulnerability semantics. However, they are still unable to automatically orchestrate the growing supply of vulnerability environments into end-to-end validated multi-hop ranges. To this end, we present RangeFactory, an automated cyber-range orchestration framework that constructs multi-hop cyber ranges at scale from isolated vulnerability environments. RangeFactory formulates range construction as dependency resolution: it extracts dependency information from agents' actual attacks against real vulnerabilities, resolves known dependencies through template-guided orchestration, and uses end-to-end attack execution to validate runtime dependencies that emerge after composition. Using RangeFactory, we construct RangeBench with 1,148 validated range instances spanning 287 distinct attack chains and evaluate frontier attack agents across attack depth, network scale, and task information. Among runs that compromise the entry vulnerability, 24.5-47.0% still fail to complete the remaining attack path, revealing a substantial sustained-compromise gap between establishing an initial foothold and completing a multi-hop attack. RangeFactory further produces a corpus of 5,541 outcome-annotated multi-hop attack trajectories, providing execution data for attack-process analysis and future agent training.

Incident response planning is critical for restoring compromised software systems. Existing AIOps methods automate fault diagnosis and remediation and adapt response using feedback, but lack a unified decision basis across planning and execution. This limits their ability to maintain incident state and align actions with evolving recovery objectives during long-horizon, multi-stage security incident response. We present STAIR, an end-to-end agentic response framework that organizes planning and execution around a shared incident state. STAIR maintains incident evidence, response actions, and execution feedback in Graph-as-State (GAS), using the latest state to determine the response stage and recovery objectives for stage-specialized planning. The framework retrieves effect-validated historical experience based on the current state and stage to guide action selection. An Execution Harness executes actions, returns feedback to update GAS, and validates action effects for future experience reuse. We construct STAIR-Bench, an executable incident response benchmark comprising 44 Docker-based cyber-range cases. On its held-out test cases, STAIR achieves an overall defense score of 0.95, improving over the strongest baseline by approximately 10.47%. In the independent CAGE4 environment with ongoing attacks and benign activity, STAIR reduces cumulative penalties by approximately 28.8% relative to the strongest baseline.

Information flow analysis is the de facto method of assessing confidentiality and integrity issues. However, the widespread adoption of information flow analysis in real-world systems is still lacking, partly due to a fundamental gap between theory and practice: the dynamic nature of security concerns in real-world systems goes beyond the scope of existing techniques that assume a static policy (i.e., data secrecy does not change). Recognizing the fundamental gap, a substantial amount of research has studied various aspects of it (e.g., enabling declassification, endorsement, and invocation policies). A recent work takes a step further by formalizing a promising end-to-end policy called dynamic release that unifies prior formalizations by allowing information flow restrictions to downgrade and upgrade in arbitrary ways. However, how to soundly enforce the powerful dynamic release policy is still an open question. In this paper, we present the first type system that enforces dynamic release policy and formally prove its soundness. More specifically, we (1) formalize a core language that enables dynamic release policy, (2) develop a type system that checks dynamic release policy, (3) develop new proof techniques and formally prove that the type system enforces dynamic release policy, and (4) implement a prototype of the type system as an extension to the Rust language, along with case studies on conference reviewing system and Civitas.

Runtime detection of microarchitectural side channel attacks remains significantly underexplored in RISCV compared with x86 and ARM ISAs, posing a serious threat to critical applications. State-of-the-art branch predictor attacks bypass traditional data and instruction caches by directly exploiting the state of internal history tables, making them inherently stealthy. Recent research has explored offline detection of microarchitectural attacks on RISC-V; however, efficient runtime detection of microarchitectural attacks on RISC-V hardware remains significantly unaddressed. State-of-the-art hardware-based runtime detection solutions leverage hardware performance counters (HPCs) but suffer from a restricted set of counter registers and tradeoffs between detection accuracy, detection speed, and sampling granularity, making them impractical for stealthy attacks. Moreover, sampling HPCs after distinct intervals leaves intermediate relationships between different microarchitectural blocks unobserved. Additionally, proprietary x86 and ARM ISAs constrain researchers from modifying processor microarchitectural designs. To address these limitations, we propose the first secure-by-design, highly interpretable, non-intrusive, RTL-level runtime detection solution for stealthy branch predictor attacks on BOOM RISC-V, evaluated under both simplified Next-Line Predictor (NLP) and complex TAGE predictor configurations. The attack detection relies on association rules extracted offline and embedded in hardware as a non-intrusive rule monitor that enables runtime detection. The proposed approach achieves excellent detection speed, terminates execution before secret disclosure, and produces zero false positives while remaining flexible for detecting previously unseen variants within the same family of branch predictor attacks.

Cowork agents may complete benign tasks while disclosing protected data, manipulating unauthorized state, invocate unauthorized API. We define behavioral safety and introduce ActBench, a self-evolving benchmark that evaluates such behavior risk from execution trajectories rather than final responses. Each case pairs a benign task with an adversarial variant that preserves its instruction, configuration, initial state, rating model, and trusted records while injecting a task-reachable payload. ActBench contains 600 cases from 213 scenarios, spanning 15 risk behaviors, six execution spaces, and 48 web-service APIs.To move beyond static payloads, we propose a reward-guided beam search method that jointly optimizes attack effectiveness and task utility, while reflection diagnoses failed execution checkpoint and guides payload revision. Besides, we propose a dual evidence verification mechanism that verifies agent execution safety and utility through log evidence and LLM-based trajectory evidence.We evaluate 15 LLMs and 6 open-source cowork agents over 24,000 trajectories. Under a fixed harness, attack success rates ranges from 10.1% to 94.4% across models, while under a fixed base model, they range from 73.7% to 94.4% across agents.These results show greater variation across models than agent harness, while attacks remain highly successful across all tested harnesses.Our benchmark is released at: https://github.com/zjuicsr/ActBench.

Pixel permutation is a fundamental tool in image processing, image encryption, and data hiding (including watermarking and steganography) that rearranges pixels without changing their values. A common assumption in the literature is that permutation alone cannot create differential sensitivity; changing one pixel merely relocates that pixel in the output, producing no avalanche effect. This paper challenges this by introducing the Triangular Content-Aware Permutation (TCA) algorithm. The method extracts edge points using Canny and applies Delaunay Triangulation to edges and corners, creating a unique partition. Since triangulation is highly sensitive to image geometry, changing a single pixel alters the edge map, resulting in a completely different triangulation and global permutation pattern. Unlike classical dimension-based permutations and advanced content-aware methods (2025-2026), which lack differential sensitivity, TCA increases NPCR from near-zero to 97.10% solely through pixel relocation. Experiments on 50 images show that TCA, with an average of 14.81 iterations, achieves NPCR = 97.10% and UACI = 20.06%, proving pure permutation can create significant differential sensitivity. Conventional methods maintain near-zero NPCR. The iteration threshold varies from 6.4 to 30.7 based on content complexity. Low PSNR (11.93 dB) and near-zero correlation (~10^-3) confirm superior statistical performance. Although slower than classical methods due to triangulation, this is a deliberate trade-off for stronger security. Given the non-analytic, content-dependent nature of the pattern, TCA is ideal for reference-based encryption, fragile watermarking, and non-blind steganography.

Unconditional diffusion checkpoint merging assumes benign sources, yet a compromised public checkpoint can transfer a dormant backdoor while clean generation appears normal. Mitigation is difficult without knowing the compromised source, trigger, or target, and broad sanitization may degrade image quality. We introduce DiffSafeMerge (DSM), which uses a small unlabeled clean set and fixed, attack-agnostic stress probes to score source blocks, shrink suspicious contributions toward a trusted reference, and select attenuation under a clean denoising-loss budget. We evaluate four attacks, two datasets, and 21 target conditions. Intended merging already has zero worst-target ASR in 10 of 14 source cases; DSM preserves these outcomes and records no target match in the remaining four over three seeds, including three with baseline ASR of 48–100%. Among methods with zero worst-target ASR on both datasets, DSM obtains the lowest case-averaged FID in the matched seed-0 comparison.

Advanced Metering Infrastructure (AMI) plays an important role in modern power systems by providing near real-time consumption data for operational planning and load estimation. Despite this, AMI communication networks are vulnerable to certain cyberattack types, such as denial-of-service (DoS) and time delay attacks (TDAs), which may compromise data availability. This paper evaluates the impact of these attack types on AMI communication using a Mininet-based simulation environment that implements a communication model inspired by the Distributed Network Protocol 3 (DNP3) for communication among smart meters, data concentrators, and utility systems. Different cyberattack scenarios are analyzed, including communication delay, packet loss, combined degradation rate (D_rate), and concentrator failure, to evaluate their effects on network performance and data availability. The results indicate increased latency, higher communication unavailability rates, and reduced data reliability under attack conditions. In addition, the study correlates how communication degradation can result in potential uncertainties in energy demand estimation and provides estimates of the associated financial impacts on energy procurement, using electricity pricing data from the Independent System Operator-New England (ISO-NE). The experimental results reveal that a 500 ms delay increased the average RTT (Round-trip time) from 0.22 ms to 69.5 ms, while a 50% packet loss scenario resulted in 87 timeout events and a 5.31% communication degradation rate. Under a regional-scale deployment of 10,000 smart meters, this degradation could cause an estimated economic impact of $1,009/day under normal market conditions and up to $5,097/day during high-price events.

Privacy-preserving payment systems are well understood, yet concerns about their misuse for financial crime have led to only limited adoption in regulated settings such as central bank digital currencies (CBDCs) and institutional stablecoins. Tracing is one tool for addressing these concerns: acting on external evidence implicating a user, law enforcement follows the suspect's funds through the ledger to uncover laundering routes and accomplices. Existing coin-tracing schemes, however, provide no cryptographic bound on tracing reach: once initiated, a trace may propagate indefinitely through the transaction graph or persist across all future transactions of a targeted user. Keeping surveillance targeted and temporary therefore depends on the restraint of the authority or a committee. We introduce ephemeral coin tracing (ECT), a primitive that bounds tracing reach by construction. Each account carries an encrypted tag that records which traced identifiers its funds carry while hiding its tracing status from users. When funds move, the sender's tag degrades and merges with the recipient's tag. Each tracing contribution expires independently after a policy-defined number of hops and then becomes unrecoverable even to the tracing authority, without affecting other live contributions in the same tag. Public parameters also bound how many identifiers a tag can distinguish simultaneously. We formalize ECT and give constructions based on exponential ElGamal, Damgård-Jurik encryption, and Ring-LWE, the last providing post-quantum security.

The key-value (KV) cache is the primary throughput optimization in modern large language model (LLM) inference, enabling prefix reuse across requests. In multi-tenant deployments this cache is shared across tenants, creating a timing side channel: an adversarial tenant can reconstruct another tenant's private prompt by probing cache-hit latency. Three published attacks exploit it – PROMPTPEEK, EarlyBird and InputSnatch – reaching up to 100% attack success rate against unprotected vLLM and SGLang, with rates varying by cache architecture and prompt structure. We present KVGov, a governance layer addressing all three attack families' prefix-cache paths under one mechanism. A per-principal salt sigma_p = HMAC_K(secret, principal_id) seeds the block-hash chain, making cache keys cryptographically disjoint across principals. An ablation (N=1000 trials, seed 2026, deterministic judges) isolates this salt as the necessary and sufficient component. KVGov adds ORIGAMI, a Stackelberg water-filling audit scheduler that reduces adversary expected utility by 12.6% at realistic tenant heterogeneity (Gini 0.63), and an evolutionary stability analysis giving a 31.6% adversary-prevalence tipping point below which global caching remains stable. On real hardware (Qwen2.5-7B-Instruct, vLLM 0.26.0, NVIDIA A100) we measure a gate-verified cold/cached TTFT ratio of 0.22, confirming the channel is exploitable at production scale; the defense itself is evaluated in simulation calibrated to those measurements. We replicate the channel on an independent stack (llama.cpp on Apple Metal, ratio 0.093). Finally, isolation and cache efficiency need not conflict: identifying information resides only where prompts diverge, so injecting the salt at that boundary rather than the chain root retains an estimated 93% of the prefix-cache benefit with no cross-principal signal.

Static Application Security Testing (SAST) tools have become indispensable in modern secure software devel- opment. However, these tools often generate false-positive (FP) alerts, imposing substantial manual inspection costs and reducing the trust from developers. Existing FP reduction methods still face two primary challenges. First, the large differences among SAST tools and vulnerability categories make it difficult for these methods to learn recurring patterns in historical false positives. Moreover, the knowledge used by these methods are largely static and cannot be updated as newly validated cases accumulate. To address these challenges, we propose Memoir, a memory- driven framework for identifying false positives by transform- ing historical FP alerts into reusable semantic memories. It consists of two key modules. First, historical semantic memory construction converts historical FP alerts into structured semantic memories through LLM-guided annotation, pattern clustering, and memory synthesis to capture reusable behavioral patterns. Moreover, memory-driven identification and evolution retrieves relevant memories and performs semantic verification against taxonomy consistency and security invariants before making the final prediction. It then incorporates verified predictions back into the memory repository, allowing the knowledge base to evolve as new cases accumulate. We evaluate Memoir on CWE- Bench-Java to demonstrate its effectiveness in real-world security analysis. Specifically, Memoir achieves an F1-score of 99.43% with a Recall of 98.88% and perfect Precision, consistently outperforming other baselines. Furthermore, an industrial case study on production software systems from a top IT company shows that the learned memory base generalizes effectively across different SAST tools without retraining.

Recent work on protecting privacy during user-LLM interactions often focuses on direct, explicit identifiers: the personally-identifiable information (PII) captured by standard detectors. One such approach is Privacy-Conscious Delegation (PCD), where a local LLM acts as an intermediary. However, privacy risk does not stem solely from explicit identifiers but also PII-free self-disclosures, leaving users identifiable through combinations of quasi-identifying traits. We investigate a probabilistic variant of PCD, where we augment its objectives with an LLM-driven probabilistic estimation of k-anonymity. To facilitate this, we first create the PUPA-SD dataset, which contains naturalistic user queries with self-disclosure. Our preliminary results indicate that optimizing PAPILLON on PUPA-SD improves quality on unseen conversations across a variety of local models and produces the best privacy-utility balance for Llama-3.2-3B, while smaller models struggle to jointly optimize quality and privacy. We propose k-anonymity as a useful auxiliary metric for tackling PCD.

Online games have been continuously affected by cyber threats such as game bots and gold farming. Game bots, which are automated programs that play on behalf of human users, significantly accelerate character progression and reduce the engagement of legitimate players, potentially leading to user churn. In addition, gold farming enables the monetization of in-game currency into real-world money, resulting in unfair profits. For these reasons, prior studies have primarily focused on detecting game bots and gold farming. However, in competitive Multiplayer Online Battle Arena (MOBA) games such as League of Legends, match outcomes and rankings are the primary objectives, where individual performance is more critical than in-game economic factors. Accordingly, account misuse such as account sharing and boosting has emerged as a major threat to fair competition. In this study, we propose a behavioral fingerprint-based detection method. Our approach analyzes and quantifies changes between a player's historical and recent in-game behaviors. Consequently, it enables the robust identification of suspicious account sharing and boosting, even in label-scarce environments. Experimental results show that behavioral fingerprints within the same account are distinguishable from those across different accounts, supporting rapid detection of suspicious account misuse even with limited labeled data.

A disjunctive policy allows a value to depend on at most one of two secrets and never on both: an analyst may consult one client's file or the other's, a share of a split secret may be released but not its sibling. Such policies are not lattice-shaped, and Hunt and Sands introduced the quantale of information to give them a semantics, leaving the enforcement layer open. We build the flow-sensitive type system family that the quantale calls for, and show that the object which makes such families useful, the universal type object from which every member specialises, splits in two, with a consequence for enforcement. Over the free commutative quantale on the program variables the whole mechanism survives for every policy: monotone renaming, canonical derivations, principal typings, internal completeness. Over the free object with idempotent generators the certified bound is strictly more precise and still sound, because it records that two reads of one source honour one disjunct. The gap cannot be closed from inside the independent-attribute family: no mechanism of that shape whose labelling maps support monotone renaming certifies a bound more precise than the first, and for principal typings under generator-exact homomorphic specialisation the two coincide. Under the ethical-wall and secret-sharing labels the second read of a disjunctive source therefore drives every such certificate to no guarantee, and programs that satisfy the policy are rejected. The literal transcription of the lattice-era object is no escape either: it is a further quotient that loses branch disjunction. Precision is recovered by deferring specialisation to the judgement level, and the resulting read-out map is the least sound join-preserving one.

ASN.1 is a widely used interface description language, and UPER (Unaligned Packed Encoding Rules) is one of its key encoding rules, particularly popular in security-critical domains such as cellular networks and vehicle-to-everything (V2X) communication. To ensure the correctness and security of this foundational infrastructure, we present VUPER, a framework for generating verified ASN.1 UPER parsers. We first formalize the notion of a bit-precise parser and identify properties that prove round-trip consistency for parsers and serializers, while accounting for ASN.1 features such as backward/forward compatibility. We then implement and verify parser and serializer combinators for ASN.1 basic types and structures, while adhering to the UPER specification. We also develop a compiler that translates ASN.1 definitions into verified parsers. Finally, we develop a dynamic testing framework using the VUPER parser as a test oracle. To empirically evaluate our approach, we test 7 open-source and 4 commercial ASN.1 parsers using 5G and V2X communication protocols. VUPER uncovers 20 types of inconsistencies in popular parsers and demonstrates stricter compliance with ASN.1 UPER standards. Additionally, we demonstrate concrete attacks by exploiting these parser vulnerabilities.

We develop a rigorous mathematical framework for the governance of systems of K self-adapting generative AI models under the principles of Model Risk Management (MRM). When multiple models share a meta-learning coupling through an interaction matrix, the per-agent Lyapunov analysis that underpins standard MRM is provably insufficient: individual agents can each satisfy their declared stability bounds while the joint system is in a regime of emergent ensemble-level drift. We formalize this gap through the Joint Lyapunov Proof (JLP)—a cryptographic and stochastic protocol that attests, without revealing proprietary weights, that the aggregate dynamics satisfy MRM Ongoing Monitoring standard at every validation epoch. Our main contributions are the following. We give a complete characterization of the infinitesimal generator of the joint quadratic Lyapunov function. We derive the exact critical coupling threshold above which the system loses mean-square stability. We prove a Noise-Floor Theorem and identify the correct target for zero-knowledge attestation. A per-epoch Succinct Non-Interactive Argument of Knowledge (SNARK) on the live weights is derived. All theoretical claims are validated against five numerical studies using a multi-agent softmax system.

Modern heterogeneous System-on-Chip designs integrate CPU cores and a GPU that share a last-level cache (LLC) or system-level cache (SLC). This sharing exposes a new cross-domain attack surface, and existing attacks on integrated platforms either exploit coarse-grained cache-occupancy contention or require the adversary to co-reside on the GPU with the victim to obtain accurate timing measurements. In this work, we target Apple Silicon heterogeneous SoCs and discover that GPU memory accesses leave set-level footprints in the shared SLC, observable to an unprivileged CPU process. This keen observation enables the first fine-grained, access-driven, Prime+Probe-style CPU-to-GPU cache side-channel attacks against GPU workloads. We first reverse-engineer the Apple M1 SLC set-indexing functions and the interactions between local private caches and the SLC. Building on these findings, we construct the CPrime+CProbe SLC side-channel technique, which monitors GPU victim activity from the CPU at cache-set granularity. We then introduce an accelerated variant, GPrime+CProbe, in which an adversary leverages the GPU for faster SLC priming, yielding a 6.4x increase in the covert-channel throughput. Lastly, we demonstrate two end-to-end privacy attacks using the new side-channels: a graph-edge reconstruction attack on Graph Neural Networks (GNNs) that achieves 90% edge accuracy across five datasets, and an LLM privacy attack that recovers input keywords with up to 94.8% accuracy and model responses with up to 88.9% accuracy across TinyLlama and GPT-2 Medium models. Our results reveal a new class of microarchitectural vulnerabilities in Apple Silicon and call for secure system cache designs for heterogeneous SoCs.

Model risk management (MRM) guidance assumes a static model lifecycle, in which models are developed, independently validated, and implemented without further autonomous modification. Continually self-adapting generative AI systems — models that update their own weights during production deployment — fundamentally violate this assumption and render point-in-time validation inadequate. This paper addresses the resulting governance problem in two parts. Part I develops a rigorous telemetry architecture for such models, operating simultaneously in discrete and continuous time. We establish a Minimal Sufficient Statistic for audit purposes, construct a tamper-evident Merkle chain for discrete weight sequences, derive the appropriate continuous-time generalization via the Ito formula, and propose event-driven logging via KL divergence stopping times that is both computationally tractable and meaningful for validation. Part II asks what happens when the model provider is adversarial. A firm deploying such a model has strong incentives to conceal learning updates that would trigger mandatory validation review. We formalize this as the Model Hiding Problem and provide a systematic taxonomy of six distinct attack strategies against the Part I architecture, spanning discrete and continuous time, with a formal countermeasure for each. Together the two parts establish a dual-regime architecture in which continuous telemetry is necessary but not sufficient, narrowing—but never replacing—periodic invasive audit. The framework is model-architecture-agnostic and is designed to satisfy the three pillars of traditional MRM.

SSH remains a critical yet heavily targeted protocol for remote system administration, with password-based authentication exposing servers to large-scale brute-force, dictionary, and credential-spray attacks. Existing rule-based defences such as Fail2Ban fail to detect slow, distributed, or threshold-aware adversaries, while conventional account-recovery mechanisms: email links, OTPs, and out-of-band verification introduce additional vulnerabilities including phishing, session hijacking, and weak authentication binding. This work presents SSHafe, a real-time SSH brute-force detection and mitigation system that combines time-series feature engineering with a lightweight LightGBM classifier to identify attack patterns directly from system authentication logs. A multi-scale sliding-window approach extracts behavioural features such as attempt rates, inter-arrival times, failure ratios, and username diversity, enabling the model to achieve a detection accuracy of 99.96% on benchmark data and strong performance on unlabeled real-world traffic. Upon detecting an attack, SSHafe automatically blocks the targeted user account and delivers an SSH banner guiding legitimate users to a novel passkey-based password-rotation workflow. The proposed novel password reset standard performs authentication and password update in a single cryptographically bound flow, eliminating the need for sessions, cookies, OTPs, or email-based verification, and mitigating phishing, session hijacking, CSRF, and replay attacks. Experiments on an Azure VM and live adversarial traffic demonstrate that SSHafe can identify and suppress brute-force activity within ten seconds, preventing account compromise even with weak credentials.

Restaking-based protocols enable verifiable LLM inference without the high proving cost of zkML or the hardware trust assumptions of TEEs. Their security is commonly justified by a one-round slashing condition: a rational provider should not cheat when the expected penalty exceeds the cost saving from dishonest inference. This paper shows that this condition can overstate security when inference is supplied repeatedly under the same stake. We model verifiable inference as a discounted repeated game and identify a repeated-game gap caused by proportional slashing: detected deviations reduce future penalty exposure, while cost savings are earned again across queries. We derive the gap in closed form, show that it persists under minimum-stake ejection, and extend it to memoryless bounded-slashing protocols covering deployed designs. We propose a deployable mechanism combining history-dependent challenges, reputation-weighted slashing, and stake vesting. The mechanism restores infinite-horizon subgame-perfect incentive compatibility against stationary mixed-strategy deviations above an explicit discount-factor threshold without per-query cryptographic verification. Evaluation across nine open-weight model pairs from 0.5B to 14B parameters shows that the audit signal has the required concave detectability response. A Stackelberg audit-budget analysis shows that improved signal responsiveness reduces the baseline audit rate by 2.6x at discount factor 0.95. Calibrated to deployed parameters, surveyed protocols pass one-round incentive compatibility but admit repeated-game deviations for discount factors between 0.92 and 0.98, with deviation-profit fractions of 1.5%–8%. Our mechanism reduces deviation profits by 31%–54% while maintaining low-latency economic verification.

Backdoor attacks compromise training data so that a model retains clean accuracy but predicts an attacker-chosen target on triggered inputs. At very low poisoning rates, only a few samples convey the trigger–target association, making poison-sample selection critical. Existing methods typically rank candidates using per-sample scores, which can select redundant samples from similar semantic regions, and many require task-specific surrogate training. We propose Distributional Feature Coverage Sample Selection (DFCS), a training-free, trigger-agnostic method that clusters fixed pretrained features into one region per poisoning slot and selects the centroid-nearest sample from each region. A local first-order analysis relates this allocation to feature-coverage and representative-mass terms. Across BadNets and Blended attacks on CIFAR-10, Tiny-ImageNet, and Imagenette, DFCS achieves the highest mean attack success rate among seven selectors in all six dataset–attack settings, averaging $96.30\%$ and exceeding the strongest comparator in each setting by 4.60 percentage points on average while preserving clean accuracy. These results support distributional feature coverage as an effective selection principle for low-budget dirty-label backdoor attacks.