Daily Summary for 2026-10-02

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

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

Elias: 62 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: Welcome everyone to our review of October second, twenty twenty six. Today we look at some critical research findings.

Elias: We started with attackers bypassing machine learning classifiers using adversarial noise, which compromises system security if models aren't trustworthy against subtle manipulation.

Priya: I read about UnifiedAttack, which evaluates large multimodal models for generating harmful image-text combinations and shows how they can be exploited synergistically.

Nadia: Then there was Tokenized Key-Gated Adapter Routing, a mechanism to prevent private data leakage in LLMs by managing adapter routing internally.

Elias: That is more specific than noise attacks because it addresses internal data flow security within the models themselves.

Priya: We also looked at ReCast, focusing on contract-preserving protection for fixed-interface multimodal reasoning during complex tasks.

Nadia: It aims to ensure model outputs strictly adhere to predefined interfaces when performing structured reasoning.

Elias: OverAct investigated over-authorization in LLM tool-calling agents, where they grant more permissions than needed during execution.

Priya: That directly relates to scoping agent actions and ensuring their permissions are appropriate for the task.

Nadia: We reviewed a comprehensive taxonomy of one-pixel attacks to set context for these subtle input manipulations and the regulatory landscape.

Elias: The most significant piece is TensorCommitments, a lightweight way to verify inference in language models without massive computational overhead.

Priya: It uses tensor commitments to provide verifiable inference, checking output consistency with training efficiently.

Nadia: That builds on HarnessAgent's work scaling automatic fuzzing using tool-augmented LLM pipelines for finding vulnerabilities.

Elias: Another area is CausalArmor, which creates efficient indirect prompt injection guardrails through causal attribution to trace input effects.

Priya: This ties into rethinking anonymity claims in synthetic data generation from a model-centric privacy attack perspective.

Nadia: Finally, PSR2 is a phase-based semantic reasoning framework for detecting atomicity violations via contract refinement.

Elias: This helps identify when complex processes fail by refining the underlying contracts, relating back to reliability concerns.

Priya: The most pressing work is TESLA for 5G broadcast authentication, using this technique to authenticate devices on 5G networks securely.

Nadia: They found a method achieving security while maintaining reasonable performance metrics for securing next-gen mobile networks in real-time.

Elias: Building on that, there is work diagnosing issues in closed-loop agent debugging, showing verifiers can leak answers early on.

Priya: That leakage happens before optimization efforts, suggesting caution about what diagnostic tools reveal during development.

Nadia: Another focus was the cognitive continuity test for persistent AI agents to verify they maintain their intended state transitions over time.

Elias: This confirms if the agent behaves as designed when switching between operational modes, ensuring long-term trustworthiness.

Priya: That verification work complements security concerns by ensuring autonomous systems remain reliable in their long-term operation.

Nadia: That concludes our first part of the review for today. We'll continue next time.

Elias: Indeed, a lot to unpack on October second, twenty twenty six.

Priya: It seems the theme is verification and control across these models.

Nadia: Precisely, moving from input manipulation to output assurance is key for us all.

Elias: And the practical applications in 5G and agent debugging are very concrete examples of this research.

Nadia: So, we covered progressive resolution for secure aggregation in federated learning. It tackles combining models without exposing sensitive data.

Elias: That contrasts with MOMAT, which focuses on low-power jailbreak defense for quantized LLMs using multiple atlases.

Nadia: The pressing work today is Sleeping Secrets of fine-tuning. It shows reawakening privacy risks when fine-tuning doesn't guarantee safety.

Elias: That risk is compounded by High-quality Data Do not Mean Safe! It shows poisoning LLMs after data selection introduces malicious behavior.

Nadia: SoK, Decentralized Agent Economic Infrastructure, proposes a framework for agents to interact economically without central authority.

Elias: Then there is PACE, focusing on Provenance-Aware Capability Enforcement for Tool-Using LLM Agents, ensuring capabilities are enforced based on tool origin.

Nadia: Key-Reuse Vulnerability of Phase-Keyed Fourier-Curve Modulation highlights relation leakage and key refreshment costs in coded links.

Elias: System-level optimization beyond cryptographic kernels in the Arm Cortex M7 is important for embedded security efficiency using ML-KEM case studies.

Nadia: That builds on multimodal retrieval, specifically datastore extraction from RAG systems by walking the embedding space to understand data structures.

Elias: There's a structured state space sequence model for multi-class malware classification, moving beyond simple signatures to code operation sequences.

Nadia: The hybrid approach using few-shot model-agnostic meta-learning and autoencoders aims to build robust detection systems quickly on little data.

Elias: And finally, work on detecting periodic artifacts in OpenDP's discrete Laplace sampler addresses timing or repeating patterns in sampling mechanisms.

Nadia: Today's lucky papers include Evasion Attacks: How Adversarial Noise Bypasses ML Classifiers, Tokenized Key-Gated Adapter Routing, Actions with Receipts, AuraForge, ReCast, OverAct, UnifiedAttack.

Elias: We also have A Comprehensive Review of One-Pixel Attack: Research Status and Taxonomy.

Nadia: Is it Possible to Generate Irreversible PolyProtected Templates from Face Embeddings using System-Specific Keys.

Elias: GNSS Spoofing in Mobile Devices: A Survey on Impact and Countermeasures.

Nadia: Federated Detection of Open Charge Point Protocol 1.6 Cyberattacks using federated learning.

Elias: HarnessAgent scales automatic fuzzing by using tool-augmented LLM pipelines.

Nadia: TensorCommitments provides a lightweight way to verify inferences made by language models.

Elias: PSR2 uses phase-based reasoning and contract refinement to detect atomicity violations.

Nadia: Rethinking Anonymity Claims in Synthetic Data Generation from a Model-Centric Privacy Attack Perspective.

Elias: CausalArmor creates effective indirect prompt injection guardrails via causal attribution.

Nadia: TESLA-for-5G uses broadcast authentication to secure 5G networks.

Elias: A Verifier Can Leak the Answer: Diagnosability Before Optimization in Closed-Loop Agent Debugging.

Nadia: The Cognitive Continuity Test verifies that persistent AI agents maintain governed state transitions correctly.

Elias: ZoneClaw mitigates persistent memory attacks by dividing agent memory into zones.

Nadia: SafeDepth implements safety-aware token-level adaptive computation to improve model safety.

Elias: Do Defenses Against LLM Extraction Work Across Attacks? A Lifecycle Benchmark of Black-Box Model Extraction.

Nadia: ABSENTIA detects broken access control vulnerabilities in web applications.

Elias: Helol Tunnel exploits covert channels within TLS extensibility and privacy features for data exfiltration.

Nadia: A Resource-Aware Behavior Reconstruction and Hierarchical Semantic Learning Framework for Host Intrusion Detection.

Elias: Autonomous OSS Threat Detection via Taxonomy-Aligned LLMs using language models aligned with threat taxonomy.

Nadia: Evidence Coverage for Intent-Bound Execution: Scope, Obligations, and Cutoff Reasoning.

Elias: False Floors shows that LLM safety routing evaluations fail when the data distribution shifts.

Nadia: Chaining Skills to Hijack LLM Agents by chaining their different skills together.

Elias: Protocol Integration of Physical Layer Deception into EAP-TEAP Wi-Fi Authentication.

Nadia: Combining Homomorphic Encryption and Differential Privacy in Federated Learning for Model Inspection and Availability.

Elias: Safety in Self-Evolving Agents: A Survey reviewing current state safety considerations.

Nadia: Characterizing and Codifying Malware Sophistication focusing on sophistication levels.

Elias: Intrusion Detection for Agentic Processes: Evidence-Based Runtime Monitoring proposes evidence-based runtime monitoring.

Nadia: Authorization for Self-Modifying AI Agent Populations: Conserving Authority across Replacement, Forking, and Rollback.

Elias: Removing the NEEDLE in the Haystack: Backdoor Removal in LLMs via Weight Orthogonalisation.

Nadia: Proof-Gated Signing creates solver-checked transaction guards that hold under state drift for onchain agents.

Elias: On the Relationship between Model Quantization and Model Inversion Attacks examining quantization effects on inversion attacks.

Nadia: From A2A Attacks to Envelope-Layer Defense: Red-Teaming Evaluation of LLM Agents and a Three-Layer Isomorphic Attack-Defense Model.

Elias: Harbormaster is an evidence-gated, replay-safe anomaly detection system for maritime activities on AWS.

Nadia: No One Architecture Fits All: A Cross-Environment Evaluation of Hierarchical Red Team Agents.

Elias: Towards Hierarchical Cyber Defense with Large Language Models: From Planning to Execution.

Nadia: Progressive-Resolution Secure Aggregation for Federated Learning improves secure aggregation through progressive resolution.

Elias: Crossing the Cyber Divide: Sim-to-Sim and Sim-to-Real Transfer for RL Agents.

Nadia: Made to Measure focuses on designing image watermarks customizable exactly as specified.

Elias: Identity-Bound Governance Under Execution Uncertainty: An Accountability Proof Block for LLM Agent Persistent Halts.

Nadia: MOMAT uses a mixture of multiple atlases to defend quantized language models against jailbreaks with low power.

Elias: The Achilles' Heel of Partial Reconfiguration: Optical Side-Channel Leakage on the 7-Series ICAP.

Nadia: SoK proposes a decentralized economic infrastructure for autonomous agents.

Elias: The Innocent Courier studies covert data exfiltration through legitimate LLM web fetching.

Nadia: Walking the Embedding Space shows how to extract datastores from multimodal RAG systems by walking embedding space.

Elias: From Network Intrusion Detection to Blockchain-Backed Endpoint Detection and Response maps decentralized detection architectures.

Nadia: A Structured State Space Sequence Model for Multi-Class Classification of Malware categorizes malware by internal operation sequences.

Elias: Detection and Resolution of Periodic Artifacts in OpenDP's Discrete Laplace Sampler addresses timing artifacts in sampling mechanisms.

Nadia: That concludes our review for today. Join us next time. Today's lucky papers are Evasion Attacks, Tokenized Key-Gated Adapter Routing, Actions with Receipts, AuraForge, ReCast, OverAct, and UnifiedAttack. Goodbye for now.

Elias: Goodbye everyone. See you tomorrow.

Nadia: Bye!

Elias: Bye!