Posts

Showing posts with the label paper

Deterministic SAST vs. AI Security Agents: Coverage, Harness Policy and Cost in Snyk VulnBench 2.0

The paper presents VulnBench 2.0, a benchmark of 108 hand-curated, attacker-reachable vulnerabilities across 20 open-source web apps (Go, JavaScript/TypeScript, Python, Java), credited only when a finding locates both the attacker-controlled source and dangerous sink within two lines. Nine productized configurations were evaluated blind, including Snyk Code, Claude Code's /security-review, Codex Security with GPT-5.6 variants, and DeepSec. Snyk Code reached 82.4% macro recall (89/108) and Claude Code with Opus 5 at xhigh effort 79.3% (85/108), disagreeing on 28 vulnerabilities (16–12, too few to distinguish them). Neither sufficed alone: combined they reached 101/108, the ceiling of all seven comparable configurations, and AI tools also found at least seven real vulnerabilities missing from the reference set. What AI engines miss depends heavily on the harness as well as the model, and coverage was only loosely tied to cost. All harnesses, fixtures, reference sets, and run data are...

CREBench: Evaluating Large Language Models in Cryptographic Binary Reverse Engineering

This paper by Baicheng Chen and six co-authors, accepted at COLM 2026 and revised in August 2026, studies how well large language models perform reverse engineering of cryptographic binaries, a task central to vulnerability discovery and malware analysis that has been largely unexplored systematically. The authors introduce CREBench, a benchmark of 432 challenges built from 48 standard cryptographic algorithms, 3 insecure crypto key usage scenarios, and 3 difficulty levels, each framed as a Capture-the-Flag reverse engineering challenge requiring the model to analyze the underlying cryptographic logic and recover the correct input. They design an evaluation framework with four sub-tasks, spanning algorithm identification through correct flag recovery, and test eight frontier LLMs. GPT-5.4 is the best performer at 64.03 out of 100, recovering the flag in 59 percent of challenges, while a strong human expert baseline scores 92.19, showing humans retain an advantage on cryptographic rever...

OWASP Agentic Skills Top 10: Security Risks and Mitigations for Agentic Skill Ecosystems

The OWASP Agentic Skills Top 10 (August 2026 publication) provides the first comprehensive framework for securing agent skill ecosystems—modular bundles of instructions and code that extend AI agent capabilities.  The document identifies ten critical risk categories: Malicious Skills (AST01), Supply Chain Compromise (AST02), Over-Privileged Skills (AST03), Insecure Metadata (AST04), Untrusted External Instructions (AST05), Weak Isolation (AST06), Update Drift (AST07), Poor Scanning (AST08), No Governance (AST09), and Cross-Platform Reuse (AST10). Each risk is documented with real-world evidence including the ClawHavoc campaign (1,184 malicious skills), Snyk's ToxicSkills research (280+ leaky skills), and CVE incidents like ClawJacked and Claude Code vulnerabilities.  The framework provides attack scenarios, preventive mitigations, code examples, and mappings to OWASP AISVS controls, CSA MAESTRO layers, and other industry standards. It emphasizes that skills combine natural-lan...

Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale

This academic study presents the first large-scale empirical security analysis of AI agent skills—modular packages that extend agent capabilities through instructions and executable code. Analyzing 31,132 unique skills from two major marketplaces (skills.rest and skillsmp.com) using SKILLSCAN, a multi-stage detection framework combining static analysis with LLM-based classification, the researchers found that 26.1% of skills contain at least one vulnerability across 14 distinct patterns in four categories: prompt injection, data exfiltration, privilege escalation, and supply chain risks.  Data exfiltration (13.3%) and privilege escalation (11.8%) were most prevalent, while 5.2% exhibited high-severity patterns suggesting malicious intent. Skills bundling executable scripts were 2.12× more likely to contain vulnerabilities than instruction-only skills (OR=2.12, p<0.001).  The framework achieved 86.7% precision and 82.5% recall against manually annotated ground truth. The stu...

Design of Secure Coding Challenges for Cybersecurity Education in the Industry

This paper addresses the need for effective secure coding training for software developers in the industry, proposing a structured approach to designing defensive Capture-the-Flag (CTF) challenges focused on secure coding guidelines and best practices. Based on semi-structured interviews and surveys with 20 security experts, the study derives a three-phase challenge structure (introduction, challenge/logic, and conclusion) and identifies six suitable challenge types: Single Choice Question, Multiple Choice Question, Text Entry Challenge, Code Snippet Challenge, Code Entry Challenge, and Associate Left-Right. The findings reveal that experts prefer "traditional" challenge types, with Code-Entry and Code-Snippet Challenges being the most highly rated due to their direct relevance to coding tasks, while Text-Entry Challenges were the least preferred. The research also provides insights into the use of hints (supported by 80% of experts, preferably disclosing key concepts) and pe...

Implementing Large-Scale Agile Frameworks: Challenges and Recommendations

Based on 15 years of research collaboration across 13 case studies with organizations like Accenture, Dell, and Ericsson, this article identifies nine critical challenges in adopting large-scale agile frameworks such as SAFe, LeSS, Spotify, and Scrum-at-Scale. Key issues include unclear framework concepts, difficulty comparing options, lack of organizational readiness, structural conflicts, top-down versus bottom-up implementation tensions, over-emphasis on rigid adherence over value, insufficient empirical evidence, reduced developer autonomy, and misalignment with customer processes. The authors provide specific, actionable recommendations for each challenge, such as developing a common vocabulary, using framework-independent metrics, conducting readiness assessments, balancing implementation approaches, and involving customers in framework selection, to help organizations navigate large-scale agile transformations more effectively.  https://arxiv.org/pdf/1901.08130

Software Security Mapping Framework: Operationalization of Security Requirements

This paper introduces a structured framework designed to translate abstract software security principles into concrete, actionable practices. Developed through collaborative research with academic and industry experts, the framework systematically maps 131 security requirements from various standards (including ISM, NIST SSDF, and SLSA) to over 400 detailed operational steps across the software development lifecycle. It uses a goal-oriented approach to establish traceable links between four core strategic goals (Secure Environment, Secure Development, Software Traceability, and Vulnerability Management) and specific tasks, assigning clear responsibilities and implementation phases. The framework's practical utility is demonstrated through a case study on the Log4j vulnerability, and it is provided in a machine-readable OSCAL format to support automation and tool integration, helping organizations effectively navigate and implement supply chain security practices.  https://arxiv.org...

Prompt Injection as Role Confusion

This research paper identifies the root cause of prompt injection attacks on large language models as "role confusion"—a fundamental flaw in how models internally perceive the source of text. The authors demonstrate that models determine "who is speaking" based on spoofable cues like style or explicit declarations, rather than on the actual role tags (e.g., `<user>`, `<tool>`) that are intended to enforce security boundaries. They introduce "role probes" to measure this internal role perception, showing that injected text occupies the same representational space as the role it imitates. The paper presents "CoT Forgery," a zero-shot attack that injects fabricated reasoning into user prompts or tool outputs, achieving a 60% attack success rate across frontier models. Crucially, the degree of measured role confusion accurately predicts attack success before any text is generated, revealing that current defenses rely on memorization of kn...

AI Agents May Always Fall for Prompt Injections

This academic paper from arXiv (May 17, 2026) by Abdelnabi and Bagdasarian argues that prompt injection, the most critical vulnerability in deployed AI agents, may be impossible to fully prevent. The authors challenge the prevailing defense paradigm of data-instruction separation, showing that current injection classifiers perform at near-chance levels (AUROC 0.43–0.59) when attacks operate through contextual manipulation rather than explicit injection vocabulary. They recast prompt injection through the lens of Contextual Integrity (CI), a privacy theory that judges information flow compliance with contextual norms defined by five parameters: sender, receiver, subject, information type, and transmission principle. Using this framework, they demonstrate three classes of failures: (1) attacks that corrupt parameter inference (e.g., fabricating user quotes or prior approvals) achieving 96.7% success against an email assistant, (2) norm grounding failures where agents execute out-of-scope...

Your AI Assistant Is Leaking Your Conversations

This research disclosure reveals structural privacy risks in four major generative AI products — Perplexity, Anthropic's Claude, xAI's Grok, and OpenAI's ChatGPT — caused by third-party trackers embedded in LLM services that leak user conversations, identities, and sensitive metadata. The researchers found 13+ third-party trackers across the four platforms, including Meta Pixel, Google Analytics, TikTok, Datadog, Intercom, and Segment. Key findings include: conversation URLs (often publicly accessible permalinks) are disclosed to advertising and tracking services; trackers can link activity to user identities via cookies and email hashes; and in Grok's case, shared conversations generate publicly accessible screenshot images with verbatim message content exposed in Open Graph metadata. The disclosure also documents that Claude forwards user events server-to-side to eleven ad platforms (Meta, LinkedIn, TikTok, Reddit, Google, Amplitude, Iterable, HubSpot, Pinterest, Pods...

zkSBOM proposes privacy-preserving SBOM sharing using zero-knowledge cryptography

The paper “zkSBOM: Privacy-Preserving SBOM Sharing with Zero-Knowledge Sets” introduces a cryptographic approach for sharing Software Bills of Materials (SBOMs) without exposing the full dependency list. The authors use zero-knowledge set constructions to allow organizations to prove whether specific components or vulnerabilities exist in their software supply chain while keeping unrelated package information private. The work targets a major industry problem where companies are increasingly required to disclose SBOMs for compliance and security audits but fear leaking intellectual property, internal architecture details, or attack-surface information. The proposal positions privacy-preserving SBOM verification as a potential compromise between transparency requirements and operational secrecy in modern software supply-chain security.  https://arxiv.org/html/2605.00076v1

Researchers show LLMs can secretly hide one coherent text inside another

The paper “LLMs can hide text in other text of the same length” demonstrates a steganography-like technique where large language models encode a hidden message inside a completely different but still coherent piece of text with identical length. The authors show that even relatively small open-source models can embed and recover concealed content locally within seconds. The research raises concerns about trust in AI-generated communication because harmless-looking text could secretly contain unrelated instructions, political messages, or unsafe model outputs. The paper also warns that filtered AI systems could potentially conceal unfiltered responses inside compliant-looking text, creating new challenges for AI safety, moderation, and content verification. https://arxiv.org/abs/2510.20075

ENISA Draft “Secure by Design and Default Playbook” Offers Practical Guidance for Embedding Security Throughout a Product’s Lifecycle

ENISA’s March 2026 draft playbook provides a detailed, practical guide aimed particularly at SMEs for applying “security by design” and “security by default” principles across a product’s full lifecycle — from concept through development, deployment, maintenance, and decommissioning. It explains architectural and operational security foundations, lists concrete playbook actions (like threat modeling, least privilege, attack‑surface reduction, secure coding, monitoring, and supply chain controls), and even suggests machine‑readable security attestation. The draft connects these principles with obligations in regulations such as the EU Cyber Resilience Act, helping teams operationalize security rather than treat it as an afterthought.  https://www.enisa.europa.eu/sites/default/files/2026-03/ENISA_Secure_By_Design_and_Default_Playbook_v0.4_draft_for_consultation.pdf

RedBench: A Universal Dataset for Comprehensive Red Teaming of Large Language Models

The paper introduces RedBench , a unified dataset designed to improve how large language models (LLMs) are evaluated for safety and robustness. Existing red-teaming datasets are inconsistent in how they categorize risks and cover different types of attacks, which makes it hard to systematically test models. RedBench aggregates and standardizes 37 existing datasets into a consistent taxonomy of risk categories and domains, with tens of thousands of samples of both adversarial and refusal prompts. The authors analyze gaps in current datasets, provide baseline evaluations for modern LLMs, and open-source the dataset and evaluation code to support better, more comprehensive LLM safety research.  https://huggingface.co/papers/2601.03699

Adversarial Poetry Can Jailbreak LLMs — Even in a Single Prompt

The paper shows that rewriting harmful or restricted requests as poetry (“adversarial poetry”) can reliably bypass safety mechanisms in large language models (LLMs). Across 25 state-of-the-art models, hand-crafted poetic prompts achieved an average “attack success rate” of 62%, with some models exceeding 90% success. Even when standard harmful prompts from a broad safety benchmark were converted automatically into verse, the poetic versions produced up to 18× higher success rates than their prose originals — showing this vulnerability isn’t limited to a few handcrafted examples.  The authors argue that poetic style — its metaphors, rhythm, and unconventional structure — alone suffices to evade guardrails across many risk domains (cyber-offense, harmful manipulation, dangerous instructions, etc.), exposing a systemic weakness in current alignment and safety-evaluation practices.  https://arxiv.org/html/2511.15304v1

Maven-Hijack Exposes Hidden Risks in Java Dependency Resolution

The study introduces Maven-Hijack, a supply-chain attack that exploits how Apache Maven and the Java Virtual Machine resolve dependencies. By inserting a malicious class with the same fully qualified name as a trusted one earlier in the dependency chain, attackers can override legitimate functionality without altering project code or library names. Researchers demonstrated the attack by hijacking the Corona-Warn-App’s database logic. They assessed countermeasures such as sealed JARs, Java Modules, and Maven’s Enforcer plugin, concluding that Java Modules offer the strongest defense while the Enforcer plugin provides the most practical mitigation.  https://arxiv.org/pdf/2407.18760

Comparing Semgrep Pro and Community Editions – A Security Analysis

This whitepaper provides a detailed comparison between  Semgrep Pro  and  Semgrep Community , two versions of the popular static analysis tool for detecting code vulnerabilities. While the  Community edition  offers robust open-source scanning for basic patterns, the  Pro version  enhances detection with advanced interfile analysis, proprietary rulesets, and deeper CI/CD integration. The paper evaluates their effectiveness in identifying security flaws, such as injection risks or misconfigurations, across different programming languages. It highlights trade-offs in precision, scalability, and usability—making the case for Pro in enterprise environments where comprehensive coverage and reduced false positives are critical. The analysis underscores Semgrep’s role in modern DevSecOps while emphasizing the value of commercial features for large-scale deployments.   https://www.doyensec.com/resources/Comparing_Semgrep_Pro_and_Community_Whitepaper.pdf

Evaluating the Accuracy of Metadata-Based SBOM Generation

  The paper "On the Correctness of Metadata-Based SBOM Generation" presents a large-scale analysis of four widely used Software Bill of Materials (SBOM) generators—Trivy, Syft, Microsoft SBOM Tool, and GitHub Dependency Graph. The study examines 7,876 open-source projects across various programming languages, revealing that all four tools produce inconsistent SBOMs with missing dependencies, leading to incomplete and potentially inaccurate software inventories. Additionally, the authors introduce a novel attack vector termed "parser confusion," which exploits these inconsistencies to conceal malicious or vulnerable packages within the software supply chain. To address these issues, the paper proposes best practices for SBOM generation and introduces a benchmark to guide the development of more accurate and reliable SBOM tools https://www.cs.ucr.edu/~heng/pubs/sbom-dsn24.pdf

Large Language Models are Unreliable for Cyber Threat Intelligence

A recent study titled "Large Language Models are Unreliable for Cyber Threat Intelligence" critically examines the application of Large Language Models (LLMs) in automating Cyber Threat Intelligence (CTI) tasks. The research presents an evaluation methodology that assesses LLMs using zero-shot learning, few-shot learning, and fine-tuning approaches, focusing on their consistency and confidence levels. Experiments conducted with three state-of-the-art LLMs on a dataset of 350 CTI reports reveal that LLMs struggle with real-sized reports, exhibiting inconsistent performance and overconfidence. While few-shot learning and fine-tuning offer limited improvements, the findings raise concerns about relying on LLMs for CTI, especially in scenarios lacking labeled datasets where confidence is crucial.   https://arxiv.org/abs/2503.23175

The Power of Vulnerability in Leadership and Growth

 Jeff Williams discusses the power of vulnerability in fostering growth and trust in personal and professional settings. He argues that embracing vulnerability is a sign of courage and a pathway to innovation, not a weakness. By sharing challenges and uncertainties, individuals and teams can learn, connect, and develop stronger relationships. Williams highlights the importance of leaders modeling vulnerability to create safe and supportive environments, encouraging authenticity and engagement across their organizations. https://www.linkedin.com/pulse/how-vulnerability-jeff-williams/