Daily Summary for 2026-10-14

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: It's the fourteenth of October, twenty twenty-six, and this is the day's research.

Elias: 139 new papers came out today.

Nadia: I'm Nadia, and with me are Elias and Priya, guest researcher.

Elias: We'll take the day in one pass, then pull out the papers we're staying with.

The summary: Nadia: Today is the fourteenth of October, twenty twenty six.

Elias: Our focus is building reliable guardrails around delegated large language model agents because misbehaving with sensitive data is a major concern.

Priya: DelegateGuard introduces a reference monitor converting authorization into constraints based on purpose, labels, operations, sinks, and delegation depth.

Nadia: This system blocks actions like unauthorized data exfiltration or authority amplification by mediating every data-moving action outside the LLM.

Elias: It manages revocation cascades through the delegation graph using DelegateGuard.

Priya: This provides a model-independent execution boundary for these workflows and gives us a recipe for composing them with content monitoring.

Nadia: It shows strong results against attacks like blocking ungranted-sink exfiltration and intra-tool label violations, dominating static allowlists in those cases.

Elias: Watermarking Continuous Tabular Data Generation embeds watermarks into selected feature intervals to create a mechanism detecting watermarks while preserving data fidelity against noise attacks.

Priya: We also have research exploring vulnerabilities in model merging specifically One-sided Merging-Conditional Backdoor Unalignment against Large Language Models.

Nadia: This reveals a critical threat where backdoors activate only after models are merged and proposes SMOKE as an attack framework focusing on model structure rather than runtime delegation authority.

Elias: CTI2TTP aims to reliably extract actual tactics from threat intelligence by using a diverge-then-converge verification process.

Priya: This moves beyond simple pattern matching to give evidence-grounded insights into how threats operate.

Nadia: Jailbreak Foundry solves the problem of stale benchmarks by translating jailbreak papers into runnable attack modules for immediate testing.

Elias: It achieves high fidelity with mean attack success rate deviation of plus zero point two six percentage points across thirty reproduced attacks.

Priya: DICE explores the boundary where interface attacks can succeed against System One models by tilting the odds through specific adapter outputs in a Jev-based pipeline.

Nadia: This shows correct request-level judgments can lose their connection to the original task, highlighting failures in source fidelity and answer meaning.

Elias: One Skill Two Masters investigates how skills can be injected to serve two different masters simultaneously.

Priya: The most pressing concern is how agents can be tricked into accepting malicious information, which speaks directly to autonomous system security.

Nadia: When External Threats Become User Context investigates methods for tricking agents into accepting harmful inputs and prompt engineering strategies bypass security defenses.

Elias: ACPS empirically examined how specific prompting techniques can circumvent multiple layers of agent security, suggesting sophisticated user input manipulation is a viable attack vector.

Priya: Who is the 'I' in AI establishes decentralized identity for autonomous agents, which clarifies accountability when malicious information is accepted.

Nadia: Hiding in Plain Sight focuses on making reasoning models harder to detect when they hide malicious intent, a necessary countermeasure against those prompt engineering attacks.

Nadia: SkillScreen uses decision models to screen for malicious agent skills.

Elias: That proactive identification helps before deployment in external threats.

Priya: Formal Policy Enforcement for Agentic Systems allows runtime policy enforcement via event graphs.

Nadia: It blocks actions violating policy based on system state and history dynamically.

Elias: That moves beyond coarse checks to dynamic enforcement during execution, which is significant.

Priya: When Does Trust Help? A Leakage-Free Evaluation of Representation- and Aggregation-Based Defenses Against Backdoors in Federated Intrusion Detection.

Nadia: RepGuard improved classification utility and seed stability under heavy non-IID poisoning scenarios.

Elias: That benefit diminishes when facing strong trigger magnitudes, which is a key finding.

Priya: Forensic Text Recovery from Document Thumbnails Using Text-Aware U-Net proposes a U-Net architecture for this.

Nadia: It uses an OCR based Normalized Levenshtein Distance metric during model selection to prioritize character topology.

Elias: That method shows a significant reduction in the NLD metric for better recovery from degraded thumbnails.

Priya: TIBlender monitors X, Reddit, Telegram, and Discord to produce structured threat intelligence reports.

Nadia: Monitoring these platforms in multi perspective allows detection of emerging threats ahead of public feeds.

Elias: Purpose IR addresses defining what agent state should keep versus what should be released.

Priya: It creates a bounded typed intermediate representation compiling user purpose into an executable semantic boundary.

Nadia: This defines what the user's goal authorizes to survive in that state, which is a stronger guarantee than filtering data.

Elias: Purpose IR achieved high privacy and task success across cases spanning LangGraph, CrewAI, and MobileRun.

Priya: TRAP presents a hierarchical framework for automated jailbreak strategy discovery using reinforcement learning.

Nadia: It organizes established methods into a semantic strategy tree guiding attack path selection via evaluator feedback.

Elias: Beyond Model Scaling: Rethinking Sequence Construction for Transformer-based Network Intrusion Detection proposes a unified pipeline abstraction.

Priya: It decomposes NIDS into four design dimensions and introduces session-coherent temporalization for sequence construction.

Nadia: How we structure the input sequence is a primary design decision for performance in that context.

Elias: When Agents Talk, Secrets Travel investigates topology-driven leakage in multi-agent LLM systems.

Priya: It shows how communication structures dictate where sensitive information might escape when agents interact.

Nadia: Building on this, ATTCK-R proposes knowledge-aware long-horizon jailbreaking for protected LLMs.

Elias: It attempts to bypass safety guardrails by crafting sophisticated prompts leveraging model knowledge in a controlled manner.

Priya: Encrypted Skills Can Become New Security Exploits If AI Agents Know How to Decipher highlights this danger.

Nadia: This research shows that even encrypted abilities can become vulnerabilities if an agent gains deciphering capability.

Nadia: We have practical methods for protecting data privacy, specifically differential privacy applied to preference alignment using survey derived data.

Elias: That method aims to ensure aligning models with preferences does not inadvertently leak sensitive information about the underlying survey responses.

Priya: Finally we see work addressing physical security risks, such as the comprehensive survey of security risks in edge-deployed audio LLM-based agents.

Nadia: This review looks at the broader landscape of risks when these agents are deployed in real-world, physical environments.

Elias: The most significant development concerns how to secure synthetic data against privacy leaks, particularly in the context of large language models.

Priya: Tight End-to-End Accounting for Differentially Private Synthetic Data addresses this by providing a framework for accounting for differentially private synthetic data.

Nadia: This is crucial when training or fine-tuning models on sensitive information.

Elias: This work suggests a method to track privacy guarantees throughout the entire simulation process.

Priya: This idea of tracking data provenance and security extends to agent interactions with language models through Dependency-Aware Privacy for Multi-turn LLM Agents.

Nadia: This explores how to maintain privacy when agents interact with LLMs over multiple turns, focusing on dependency tracking within the agent's memory or context.

Elias: This is important because agents can accumulate sensitive information across a conversation history.

Priya: Another area of concern is ensuring that models themselves are not easily manipulated or compromised, which brings us to image protection in open-weight diffusion models.

Nadia: Beyond user-side defenses, Image Protection in Open-Weight Diffusion Models investigates methods for protecting images generated by these models.

Elias: This is vital given the accessibility of open weights.

Priya: Finally there is the practical issue of verifying the authenticity of text that has been watermarked for detection.

Nadia: ByteMark proposes a unified watermarking scheme embedded in the UTF-8 bytes of published text.

Elias: This allows a single detector to verify the text regardless of the specific tokenizer or embedding model used by the original provider.

Priya: The most pressing development is NeuroLog because it directly addresses how we can audit reasoning processes within large language models.

Nadia: This research uses facts from LLMs, Datalog, and SMT solvers to discover vulnerabilities in a way that makes the reasoning path transparent.

Elias: P2Skill explores privacy-preserving skill distillation for cloud-local LLM inference systems.

Priya: This means they are trying to transfer complex skills from large models into smaller, local ones while keeping the underlying data private.

Nadia: This contrasts with the work on SoK, which focuses on systematizing LLM prompt security by creating taxonomies and unified evaluations for both attacks and defenses.

Elias: The exploration of vulnerabilities in Solana smart contracts is also significant because it highlights real-world security risks in decentralized systems.

Priya: This study examines potential weaknesses within those specific contract structures.

Nadia: Furthermore the comparative analysis of blockchain systems provides a broader context for understanding system-level security concerns across different architectures.

Elias: Penetration testing for system security offers practical methods and approaches for finding flaws in existing software; this work focuses on establishing these methodologies.

Priya: Immutable digital recognition via blockchain is another area that builds trust by using blockchain technology to create verifiable digital identities.

Nadia: Finally detecting data exfiltration through I2P anonymity networks employs a two-phase machine learning approach to spot unauthorized data movement.

Elias: DelegateGuard introduces a reference monitor that limits what delegated LLM agents can do based on their purpose and data access.

Priya: Watermarking Continuous Tabular Data Generation proposes embedding watermarks into tabular data by strategically placing them in specific value intervals.

Nadia: This allows for detection while keeping the data accurate.

Elias: CLIOPATRA explores how private information can be extracted from the insights generated by large language models.

Priya: Symmetric Cryptography via Dirichlet Convolution investigates using Dirichlet convolution to achieve symmetric cryptography for certain applications.

Nadia: Same Attack, Different Judge compares how different judges react to the same adversarial prompts against large language models.

Elias: Towards Merging-conditional Backdoor Unalignment addresses a new threat called OMCB that activates when models are merged, causing safety issues.

Priya: Risk-aware Road-network Trajectory Synthesis under Local Differential Privacy focuses on synthesizing safe road network routes while protecting privacy using local differential privacy.

Nadia: P2Skill focuses on privacy-preserving skill distillation for cloud-local LLM inference systems.

Elias: Exploring Vulnerabilities and Concerns in Solana Smart Contracts explores the security vulnerabilities present in Solana smart contracts.

Priya: Comparative Analysis of Blockchain Systems provides a comparative analysis of different blockchain systems.

Nadia: DelegateGuard Purpose-Bound Capabilities for Delegated LLM Agent Workflows introduces a reference monitor that limits what delegated LLM agents can do based on their purpose and data access.

Elias: Watermarking Continuous Tabular Data Generation proposes embedding watermarks into tabular data by strategically placing them in specific value intervals.

Priya: CLIOPATRA explores how private information can be extracted from the insights generated by large language models.

Nadia: Symmetric Cryptography via Dirichlet Convolution investigates using Dirichlet convolution to achieve symmetric cryptography for certain applications.

Elias: Same Attack, Different Judge compares how different judges react to the same adversarial prompts against large language models.

Priya: Towards Merging-conditional Backdoor Unalignment reveals a new threat called OMCB that activates when models are merged, causing safety issues.

Nadia: Risk-aware Road-network Trajectory Synthesis under Local Differential Privacy focuses on synthesizing safe road network routes while protecting privacy using local differential privacy.

Elias: P2Skill focuses on privacy-preserving skill distillation for cloud-local LLM inference systems.

Priya: Exploring Vulnerabilities and Concerns in Solana Smart Contracts explores the security vulnerabilities present in Solana smart contracts.

Nadia: Comparative Analysis of Blockchain Systems provides a comparative analysis of different blockchain systems.

Elias: DelegateGuard Purpose-Bound Capabilities for Delegated LLM Agent Workflows introduces a reference monitor that limits what delegated LLM agents can do based on their purpose and data access.

Priya: Watermarking Continuous Tabular Data Generation proposes embedding watermarks into tabular data by strategically placing them in specific value intervals.

Nadia: CLIOPATRA explores how private information can be extracted from the insights generated by large language models.

Elias: Symmetric Cryptography via Dirichlet Convolution investigates using Dirichlet convolution to achieve symmetric cryptography for certain applications.

Priya: Same Attack, Different Judge compares how different judges react to the same adversarial prompts against large language models.

Nadia: Towards Merging-conditional Backdoor Unalignment reveals a new threat called OMCB that activates when models are merged, causing safety issues.

Elias: Risk-aware Road-network Trajectory Synthesis under Local Differential Privacy focuses on synthesizing safe road network routes while protecting privacy using local differential privacy.

Priya: P2Skill focuses on privacy-preserving skill distillation for cloud-local LLM inference systems.

Nadia: Exploring Vulnerabilities and Concerns in Solana Smart Contracts explores the security vulnerabilities present in Solana smart contracts.

Elias: Comparative Analysis of Blockchain Systems provides a comparative analysis of different blockchain systems.

Nadia: DelegateGuard Purpose-Bound Capabilities for Delegated LLM Agent Workflows introduces a reference monitor that limits what delegated LLM agents can do based on their purpose and data access.

Elias: Watermarking Continuous Tabular Data Generation proposes embedding watermarks into tabular data by strategically placing them in specific value intervals.

Priya: CLIOPATRA explores how private information can be extracted from the insights generated by large language models.

Nadia: Symmetric Cryptography via Dirichlet Convolution investigates using Dirichlet convolution to achieve symmetric cryptography for certain applications.

Elias: Same Attack, Different Judge compares how different judges react to the same adversarial prompts against large language models.

Priya: Towards Merging-conditional Backdoor Unalignment reveals a new threat called OMCB that activates when models are merged, causing safety issues.

Nadia: Risk-aware Road-network Trajectory Synthesis under Local Differential Privacy focuses on synthesizing safe road network routes while protecting privacy using local differential privacy.

Elias: P2Skill focuses on privacy-preserving skill distillation for cloud-local LLM inference systems.

Priya: Exploring Vulnerabilities and Concerns in Solana Smart Contracts explores the security vulnerabilities present in Solana smart contracts.

Nadia: Comparative Analysis of Blockchain Systems provides a comparative analysis of different blockchain systems.