Searcharxiv⌕ Search

arXiv subjects

Dinil Mon Divakaran

Publications and source records attributed to Dinil Mon Divakaran.

At least 19 recordsLinked to original sources

The Truth Beneath the Tokens: Prompt-Space Recovery of Knowledge Edits in Language Models

Model editing provides an efficient way to update factual knowledge in transformer models. Recent approaches reverse edits by modifying, masking, or decomposing model weights, but make recovery dependent on weight-space interventions. We instead show that pre-edit knowledge can be recovered through input-space interventions alone, without modifying the edited model's weights. We introduce Memento, a prompt-space mechanism for auditing residual knowledge that learns a reusable recovery signal to steer an edited model toward its pre-edit response. Memento supports both continuous recovery prefixes and discrete textual prompts for standard text interfaces. Across three LLMs and five editing techniques, Memento recovers 80.07% of held-out edited facts, achieving a 9.58% relative improvement over baselines, suggesting that editing may suppress the facts without fully eliminating their recoverability. We further investigate whether prompt-space recovery extends beyond model editing. On pristine, unedited LLMs, Memento improves average factual recall by 53.9% relative to standard inference. These results demonstrate the potential of prompt-space recovery for auditing residual knowledge in edited models and investigating the gap between factual knowledge and its accessibility during inference. We will release our source code.

cs.CL↗

CLaRE-ty Amid Chaos: Quantifying Representational Entanglement to Predict Ripple Effects in LLM Editing

The static knowledge representations of large language models (LLMs) inevitably become outdated or incorrect over time. While model-editing techniques offer a promising solution by modifying a model's factual associations, they often produce unpredictable ripple effects, which are unintended behavioral changes that propagate even to the hidden space. In this work, we introduce CLaRE, a lightweight representation-level technique to identify where these ripple effects may occur. Unlike prior gradient-based methods, CLaRE quantifies entanglement between facts using forward activations from a single intermediate layer, avoiding costly backward passes. To enable systematic study, we prepare and analyse a corpus of 11,427 facts drawn from three existing datasets. Using CLaRE, we compute large-scale entanglement graphs of this corpus for multiple models, capturing how local edits propagate through representational space. These graphs enable stronger preservation sets for model editing, audit trails, efficient red-teaming, and scalable post-edit evaluation. In comparison to baselines, CLaRE achieves an average of 62.2% improvement in Spearman correlation with ripple effects while being 2.7x faster, and using 2.85x less peak GPU memory. Besides, CLaRE requires only a fraction of the storage needed by the baselines to compute and preserve fact representations. Our entanglement graphs and corpus are available at https://github.com/manitbaser/CLaRE.

cs.LG↗

NetGuise: Adversarial Traffic Generation for Early-Stage Classifier Evasion

Early-stage encrypted traffic classifiers make decisions from only the first few packets of a connection, enabling low-latency classification but raising robustness concerns under realistic traffic morphing. Existing adversarial traffic generation methods often assume complete flows, feature-level perturbations, or iterative runtime feedback, which do not match the constraints of early-stage execution. We present NetGuise, a target-conditioned traffic morphing framework that generates packet-level padding and non-negative delay actions while preserving packet direction, ordering, and packet count and respecting MTU, bandwidth, and latency constraints. We evaluate NetGuise for censorship evasion under two observation interfaces, i) a history-only causal interface for online packet-by-packet execution, ii) and a full-window oracle interface that provides an upper-bound characterization of early-stage morphability. On VANT, NetGuise achieves 33.2% mean targeted evasion under the causal interface versus 82.0% under the oracle interface; on UTMobileNet, the corresponding results are 5.6% and 81.8%. NetGuise outperforms the strongest causal baseline, BAP, on both datasets and, on VANT, reduces bandwidth overhead from 19.80% to 3.02% and generation latency from 37.42 ms/flow to 1.16 ms/flow. Under the oracle interface, it achieves up to 85.7% targeted evasion, retains 43.5% after four rounds of adaptive censor retraining, and maintains 48% under zero-shot cross-environment transfer. Our results show that early-stage classifiers can remain morphable despite high clean accuracy and that robustness evaluation should account for targeted morphability, observation constraints, defender adaptation, and operational network conditions.

cs.NI↗

ExeBench-FHD: Addressing Fidelity Issues in Decompiler Evaluation

ExeBench is one of the most widely used benchmarks for evaluating LLM-based decompilers. We analyze ExeBench and identify nine issues that affect how faithfully a test measures a decompiler. Three of them prevent LLM decompilers from recovering C code from ExeBench's function. These are ambiguous parameter order when integer and floating-point types are mixed, unused parameters that optimized assembly may drop, and constants stored outside the function's assembly code. Six more affect test validity: non-void functions without a return, placeholder types, static or inline functions removed by the optimizer, functions named main, trivial bodies and truncated definitions. We present ExeBench-FHD, a backward-compatible revision of ExeBench that adds 16 new fields: call-site assembly, data sections, test wrappers, linked-binary disassembly and five warning flags. In ExeBench, 38.9% of functions have at least one of the five flags; dropping records with any of the four warning flags (all except unused_param, which marks functions that the new wrappers make evaluable) keeps 63.9% of the corpus. In an evaluation at -O2 on four architectures, the new fields raise GPT 6 Sol's re-execution rate by up to 26.5%, 19.1% and 21.0% on functions that read data sections, mix parameter types and have unused parameters, respectively.

cs.SE↗

DarKnight: Retrieval-Guided Agentic Framework for Breaking Dark Web CAPTCHA Gates

Cyber Threat Intelligence (CTI) investigators who monitor dark web marketplaces for leaked credentials and breach listings must pass the diverse CAPTCHAs that operators deploy to block automated collection. Existing solvers either need unscalable per-site fine-tuning or rely on commercial Vision-Language Model (VLM) APIs, which surrender operational sovereignty: provider policies, Tor-exit blocking, government regulation, or model updates outside the investigator's control can halt or alter the solver overnight, and per-call costs recur with every session and retry. We present DarKnight, a self-improving agentic framework designed to run on a single workstation GPU inside an investigator's Tor crawler and solve CAPTCHAs with small, locally hosted open-weight VLMs that no provider decision can withdraw, without per-market training or third-party services. For each CAPTCHA, DarKnight retrieves successful and failed past attempts on similar images (dual-signal RAG), composes an image-specific preprocessing pipeline, and learns from the site's pass/fail verdict. In a case study on four real-world dark web marketplaces, agentic tool selection outperformed parameter scaling. A deployable open-weight VLM with DarKnight reached 83.2\% on the most heavily defended market, 37.2 percentage points above its base model, outperforming even the frontier Claude Opus~5 (44%). DarKnight recovered 48.2% of baseline failures at a 7.6% degeneration cost, indicating that selective tool use can avoid much of the degradation caused by uniform preprocessing.

cs.CR↗

SoK: Are LLMs Reliable at Source Code Recovery? A Taxonomy and Empirical Evaluation

Effective source recovery is critical to security applications such as malware analysis, vulnerability assessment, and legacy maintenance. Large Language Models (LLMs) are reshaping this field, shifting the paradigm away from rule-based heuristics to probabilistic and high fidelity semantic recovery of source code from assembly or classical decompiler-derived pseudo-C. However, despite rapid progress, the field suffers from fragmentation across numerous approaches as well as their non-unified evaluations, limiting objective comparisons. Further, existing works have limited coverage of embedded, IoT architectures and source languages beyond C/C++. In this work, we present the first Systematization of Knowledge (SoK) focused specifically on LLM-assisted binary-to-source recovery. We provide a granular design-centric taxonomy of LLM-assisted source recovery methods and systematic evaluations using seven key metrics along six evaluation dimensions. We evaluate state-of-the-art methods on 45,000 test samples derived from four standard and five embedded architectures, five optimization levels, and symbol stripping. We ablate the impact of design choices on recovery performance, including input representation, contextual enrichment, model scale, iteration and review roles using controlled in-house recovery pipelines and three off-the-shelf models. Finally, we test same-language and cross-language recovery capability covering four mature and two legacy languages. Our systematization and comprehensive evaluations provide key insights that guide future directions in this field.

cs.CR↗

The Fragility of Trigger-Tag Mechanisms for Misuse Detection in Open-Weight LLMs

Open-weight language models can be downloaded, modified, and deployed beyond their developers' control, limiting the effectiveness of centrally enforced safeguards. Recent work has therefore proposed \emph{trigger-tag} mechanisms that produce a detectable signal when a model is used under a target condition, such as generating phishing contents. Although these mechanisms borrow from established techniques, their use for conditional misuse detection in open-weight LLMs is relatively new. Therefore, existing research works have not systematically studied the robustness of trigger-tag mechanisms under adversarial attacks. To close this gap, (i)~we formalize trigger-tags and distinguish \emph{token-level trigger-tags}, which introduce watermark-inspired signals during decoding, from \emph{weight-level trigger-tags}, which learn backdoor-inspired associations between target conditions and detectable model behavior. Furthermore, (ii)~we introduce \Untag, a unified attack framework that organizes their mechanism-specific attack surfaces into a common taxonomy. We evaluate representative token-level and weight-level trigger-tags using phishing as a case study. We find that while trigger-tags may provide useful evidence in controlled settings, our attacks render the existing trigger-tag mechanisms to be entirely ineffective. Consequently, we argue that these mechanisms should not be treated as robust misuse detectors when attackers can transform outputs or modify open weights.

cs.CR↗

The Tokens Remember: When Tokenization Bypasses Knowledge Editing and Unlearning

Open-weight LLMs give downstream users control over the inference stack, but this flexibility can undermine post-release guarantees that sensitive knowledge has been modified or removed. Model editing and machine unlearning are used to modify or remove targeted knowledge without retraining models from scratch. However, existing security evaluations of these techniques face two critical limitations. First, they typically require access to either the original pre-edit/unlearning model or auxiliary classifiers to detect modifications or reconstruct pre-edit behavior. Second, they evaluate modifications under the canonical tokenization of an input, implicitly treating tokenization as a benign preprocessing step. We show that this assumption creates a security gap: the same input string can be represented by alternative valid tokenizations that induce different computational trajectories, allowing an adversary to bypass localized modifications and recover information intended to be suppressed. We introduce Toketive, a simple yet powerful reference-free attack that exploits the tokenization-based side channel to (i) detect modified knowledge and (ii) reconstruct the corresponding pre-edit response. It operates solely on the released model and requires neither the pre-edit model, training data, shadow models, nor auxiliary classifiers. Across five LLMs, six datasets, and six editing and unlearning techniques, we find that 38.6% of alternative tokenizations bypass the modification and recover the pre-edit response. Toketive detects modified facts with an F1 score of 84.2%, a 26.2% relative gain over the strongest baseline, and reconstructs pre-edit responses with 74.5% top-5 accuracy, 21.7% higher than the best baseline. Our results show that localized modifications should not be treated as robust knowledge-control boundaries without adversarial evaluation over alternative representations.

cs.CR↗

Drishti: AI-Led Human-Directed Vulnerability Auditing for 5G Cores

Candidate generation for open-source vulnerabilities is no longer scarce. AI-assisted code review now produces defect candidates cheaply, and industry programs pair them with expert human triage. The remaining scarcity is validation and impact assessment, and the gap is largest in critical-infrastructure software like 5G cores. Here, validation has four costs: verification, reachability, impact, and fix-completeness. We present Drishti, an AI-led human-directed vulnerability audit framework with four components, one per cost: (i) an anti-pattern catalog for verification, (ii) critical-path triage for reachability, (iii) concentric validation for impact, and (iv) patch-review for fix-completeness. Across audits of Open5GS and free5GC, Drishti produced three findings. The first is a pre-authentication NULL-dereference in the Open5GS NRF multipart parser, fixed upstream with a CVE requested. The second is an ASN.1-PER memory amplification in the free5GC NGAP decoder. A 2-byte input from a rogue gNodeB OOM-kills the AMF in 6.2 seconds. The third is a defective patch on CVE-2025-69248 whose defense-in-depth check is dead code before authentication.

cs.CR↗

CHISEL-ing Back Source Code with AI-enabled Iterative Recovery

Decompilation aims to recover high-level, compilable, and semantically equivalent code from binaries. Traditional decompilers produce pseudo-C that is difficult to read and does not compile, while the recent LLM-assisted approaches generate readable, but semantically incorrect code. LLM-aided iterative recovery is an emerging branch of research, but prior works rely on supplied test suites for semantic recovery. In this work, we present CHISEL, a test suite-free framework to iteratively recover source code from Ghidra-derived pseudo-C. CHISEL uses simple yet effective feedback from a compiler (static analysis) and a coverage-guided fuzzer (differential analysis), augmented by rich observables for grounded divergence detection and feedback, cross-iteration divergence memory, and best candidate retention. We systematically evaluate CHISEL for compilation and semantic recovery, feedback oracle soundness, and iteration overhead on 120 ExeBench functions compiled for the x86-64 architecture, across four optimizations (O0-O3), in both stripped and unstripped variants, using the open-weight Gemma4:31b LLM. CHISEL, with all recommended features, achieves an average of 96.1% re-compilability and 79.8% re-executability rates at an average of 2.1 iterations. Significantly, CHISEL recovers 26% of first-generation execution errors. At the same time, CHISEL feedback oracle falsely accepts only 9.4% candidates. Lastly, CHISEL performs significantly better than two recent prior work on LLM-assisted decompilation.

cs.CR↗

MATRA: Modeling the Attack Surface of Agentic AI Systems -- OpenClaw Case Study

LLMs are increasingly deployed as autonomous agents with access to tools, databases, and external services, yet practitioners (across different sectors) lack systematic methods to assess how known threat classes translate into concrete risks within a specific agentic deployment. We present MATRA, a pragmatic threat modeling framework for agentic AI systems that adapts established risk assessment methodology to systematically assess how known LLM threats translate into deployment-specific risks. MATRA begins with an asset-based impact assessment and utilizes attack trees to determine the likelihood of these impacts occurring within the system architecture. We demonstrate MATRA on a personal AI agent deployment using OpenClaw, quantifying how architectural controls such as network sandboxing and least-privilege access reduce risk by limiting the blast radius of successful injections.

cs.AI↗

ThinkEval: Practical Evaluation of Knowledge Leakage in LLM Editing using Thought-based Knowledge Graphs

Robust model-editing techniques are essential for deploying large language models (LLMs) in practical applications, as they enable cost-effective ways to deal with challenges such as privacy breaches, bias mitigation and misinformation spread. For example, an LLM-based healthcare assistance may need to update out-dated or incorrect knowledge to prevent harmful recommendations. However, many editing techniques focus on isolated facts, which critically fail to prevent indirect knowledge leakage -- the unintended reconstruction of edited-out information through persistent causal links and contextual relationships. To assist users in selecting the right editing technique, we develop and present ThinkEval, a framework to systematically quantify indirect knowledge leakage and ripple effects in model-editing. ThinkEval builds and employs specialized knowledge graphs to analyze the causal structure of facts before and after editing. To support this approach, we present KnowGIC, a benchmark dataset comprising multi-step reasoning paths that precisely measure these complex knowledge transformation effects. We evaluate five editing techniques: AlphaEdit, RECT, ROME, MEMIT, and PRUNE across multiple LLMs. Our results show that these techniques struggle to balance indirect fact suppression with the preservation of related knowledge, compromising the contextual integrity of a model's knowledge. Our dataset is available at: https://github.com/manitbaser/KnowGIC.

cs.LG↗

AI-based Traffic Modeling for Network Security and Privacy: Challenges Ahead

Network traffic analysis using AI (machine learning and deep learning) models made significant progress over the past decades. Traffic analysis addresses various challenging problems in network security, ranging from detection of anomalies and attacks to countering of Internet censorship. AI models are also developed to expose user privacy risks as demonstrated by the research works on fingerprinting of user-visiting websites, IoT devices, and different applications, even when payloads are encrypted. Despite these advancements, significant challenges remain in the domain of network traffic analysis to effectively secure our networks from evolving threats and attacks. After briefly reviewing the relevant tasks and recent AI models for traffic analysis, we discuss the challenges that lie ahead.

cs.CR↗

RECTor: Robust and Efficient Correlation Attack on Tor

Tor is a widely used anonymity network that conceals user identities by routing traffic through encrypted relays, yet it remains vulnerable to traffic correlation attacks that deanonymize users by matching patterns in ingress and egress traffic. However, existing correlation methods suffer from two major limitations: limited robustness to noise and partial observations, and poor scalability due to computationally expensive pairwise matching. To address these challenges, we propose RECTor, a machine learning-based framework for traffic correlation under realistic conditions. RECTor employs attention-based Multiple Instance Learning (MIL) and GRU-based temporal encoding to extract robust flow representations, even when traffic data is incomplete or obfuscated. These embeddings are mapped into a shared space via a Siamese network and efficiently matched using approximate nearest neighbor (aNN) search. Empirical evaluations show that RECTor outperforms state-of-the-art baselines such as DeepCorr, DeepCOFFEA, and FlowTracker, achieving up to 60% higher true positive rates under high-noise conditions and reducing training and inference time by over 50%. Moreover, RECTor demonstrates strong scalability: inference cost grows near-linearly as the number of flows increases. These findings reveal critical vulnerabilities in Tor's anonymity model and highlight the need for advanced model-aware defenses.

cs.CR↗

UniNet: A Unified Multi-granular Traffic Modeling Framework for Network Security

As modern networks grow increasingly complex--driven by diverse devices, encrypted protocols, and evolving threats--network traffic analysis has become critically important. Existing machine learning models often rely only on a single representation of packets or flows, limiting their ability to capture the contextual relationships essential for robust analysis. Furthermore, task-specific architectures for supervised, semi-supervised, and unsupervised learning lead to inefficiencies in adapting to varying data formats and security tasks. To address these gaps, we propose UniNet, a unified framework that introduces a novel multi-granular traffic representation (T-Matrix), integrating session, flow, and packet-level features to provide comprehensive contextual information. Combined with T-Attent, a lightweight attention-based model, UniNet efficiently learns latent embeddings for diverse security tasks. Extensive evaluations across four key network security and privacy problems--anomaly detection, attack classification, IoT device identification, and encrypted website fingerprinting--demonstrate UniNet's significant performance gain over state-of-the-art methods, achieving higher accuracy, lower false positive rates, and improved scalability. By addressing the limitations of single-level models and unifying traffic analysis paradigms, UniNet sets a new benchmark for modern network security.

cs.CR↗

Step-by-Step Reasoning Attack: Revealing 'Erased' Knowledge in Large Language Models

Knowledge erasure in large language models (LLMs) is important for ensuring compliance with data and AI regulations, safeguarding user privacy, mitigating bias, and misinformation. Existing unlearning methods aim to make the process of knowledge erasure more efficient and effective by removing specific knowledge while preserving overall model performance, especially for retained information. However, it has been observed that the unlearning techniques tend to suppress and leave the knowledge beneath the surface, thus making it retrievable with the right prompts. In this work, we demonstrate that \textit{step-by-step reasoning} can serve as a backdoor to recover this hidden information. We introduce a step-by-step reasoning-based black-box attack, Sleek, that systematically exposes unlearning failures. We employ a structured attack framework with three core components: (1) an adversarial prompt generation strategy leveraging step-by-step reasoning built from LLM-generated queries, (2) an attack mechanism that successfully recalls erased content, and exposes unfair suppression of knowledge intended for retention and (3) a categorization of prompts as direct, indirect, and implied, to identify which query types most effectively exploit unlearning weaknesses. Through extensive evaluations on four state-of-the-art unlearning techniques and two widely used LLMs, we show that existing approaches fail to ensure reliable knowledge removal. Of the generated adversarial prompts, 62.5% successfully retrieved forgotten Harry Potter facts from WHP-unlearned Llama, while 50% exposed unfair suppression of retained knowledge. Our work highlights the persistent risks of information leakage, emphasizing the need for more robust unlearning strategies for erasure.

cs.CR↗

From ML to LLM: Evaluating the Robustness of Phishing Webpage Detection Models against Adversarial Attacks

Phishing attacks attempt to deceive users into stealing sensitive information, posing a significant cybersecurity threat. Advances in machine learning (ML) and deep learning (DL) have led to the development of numerous phishing webpage detection solutions, but these models remain vulnerable to adversarial attacks. Evaluating their robustness against adversarial phishing webpages is essential. Existing tools contain datasets of pre-designed phishing webpages for a limited number of brands, and lack diversity in phishing features. To address these challenges, we develop PhishOracle, a tool that generates adversarial phishing webpages by embedding diverse phishing features into legitimate webpages. We evaluate the robustness of three existing task-specific models - Stack model, VisualPhishNet, and Phishpedia - against PhishOracle-generated adversarial phishing webpages and observe a significant drop in their detection rates. In contrast, a multimodal large language model (MLLM)-based phishing detector demonstrates stronger robustness against these adversarial attacks but still is prone to evasion. Our findings highlight the vulnerability of phishing detection models to adversarial attacks, emphasizing the need for more robust detection approaches. Furthermore, we conduct a user study to evaluate whether PhishOracle-generated adversarial phishing webpages can deceive users. The results show that many of these phishing webpages evade not only existing detection models but also users.

cs.CR↗

EagleEye: Attention to Unveil Malicious Event Sequences from Provenance Graphs

Securing endpoints is challenging due to the evolving nature of threats and attacks. With endpoint logging systems becoming mature, provenance-graph representations enable the creation of sophisticated behavior rules. However, adapting to the pace of emerging attacks is not scalable with rules. This led to the development of ML models capable of learning from endpoint logs. However, there are still open challenges: i) malicious patterns of malware are spread across long sequences of events, and ii) ML classification results are not interpretable. To address these issues, we develop and present EagleEye, a novel system that i) uses rich features from provenance graphs for behavior event representation, including command-line embeddings, ii) extracts long sequences of events and learns event embeddings, and iii) trains a lightweight Transformer model to classify behavior sequences as malicious or not. We evaluate and compare EagleEye against state-of-the-art baselines on two datasets, namely a new real-world dataset from a corporate environment, and the public DARPA dataset. On the DARPA dataset, at a false-positive rate of 1%, EagleEye detects $\approx$89% of all malicious behavior, outperforming two state-of-the-art solutions by an absolute margin of 38.5%. Furthermore, we show that the Transformer's attention mechanism can be leveraged to highlight the most suspicious events in a long sequence, thereby providing interpretation of malware alerts.

cs.CR↗