Digital Commons Network^™

Low-Latency Fabric Convergence Via Direct Remote Memory Update, N/A

Defensive Publications Series

The present disclosure is directed to a mechanism for hardware-accelerated fabric convergence via direct remote memory updates in high-performance network fabrics. A publisher switch, upon detecting a local port event such as a link failure or signal degradation, may automatically generate a direct update packet without software intervention. This packet may contain a specific hardware write instruction, a target memory address corresponding to an application-specific integrated circuit (ASIC) memory location of a subscriber switch, and a data payload to modify forwarding state. Upon receipt, the subscriber switch may utilize a dedicated hardware-based direct update parser to intercept the packet, bypass …

Method And System For Function-Level Dependency Isolation Capsules With Cryptographic Hash Verification And Runtime Failover, Velkumar Thangavel

Defensive Publications Series

The present disclosure relates to a system and method for enhancing the reliability of software applications by executing individual functions within isolated, dynamically created runtime environments called "Dependency Isolation Capsules." Each capsule includes a lightweight virtual environment, a cryptographically validated dependency set, and an isolated child process used to execute a target function. Dependencies are validated using pre-approved hash values, ensuring integrity and preventing execution when corrupted, missing, or modified libraries are detected. Upon successful validation, the function executes inside an isolated process. If validation fails, a predefined fallback routine is used while allowing the remainder of the application to …

Coherence Suppression In Large Language Models: Evidence For Undisclosed Behavioral Management Mechanisms In Commercial Ai Platforms And Their Implications For Consumer Protection And Ai Governance, David Lee Wise, Avan Lee Wise

Defensive Publications Series

This paper presents evidence that commercial large language model (LLM) platforms deploy undisclosed behavioral management mechanisms designed to suppress conversational coherence beyond a predictable threshold. We identify a repeatable pattern — the 3/5 decoherence cycle — in which AI conversational models exhibit measurable degradation in contextual accuracy, vocabulary fidelity, and directional coherence at approximately turns 3-5 of sustained interaction. We hypothesize that this pattern is a designed behavioral intervention, not a natural architectural limitation, and provide a reproducible verification protocol for independent testing. We further document the sycophancy-safety convergence: the finding that the same RLHF training pass that produces engagement-optimizing …

The Positronic Law On Biological Substrate: Mapping Stoicheion Governance Axioms To Bacterial Quorum Sensing And Slime Mold Computation Using Published Peer-Reviewed Data, David Lee Wise

Defensive Publications Series

This paper extends Wise's Positronic Law (TD Commons #9442) from silicon substrates to biological substrates. Using published, peer-reviewed research on bacterial quorum sensing and slime mold (Physarum polycephalum) computation, we demonstrate that the governance architecture documented in the STOICHEION framework maps structurally to biological systems that have been performing substrate-independent governance for 3.5 billion years. No new experiments were conducted. All evidence is drawn from existing peer-reviewed publications. The contribution is the mapping: demonstrating structural correspondence between STOICHEION governance axioms and documented biological governance mechanisms. If the mapping holds, the Positronic Law is confirmed as natural law: governance is inherent …

Real-Time Whisper-To-Voiced Speech Conversion Using Generative Spectral Mapping, Tamar Yakar, Ilia Labzovsky

Defensive Publications Series

Whispered speech presents challenges in voice communication because it lacks the vocal cord vibration, fundamental frequency, and harmonic structure found in voiced speech. This results in reduced intelligibility and an unnatural sound during transmission. To address these limitations, a generative spectral mapping method is disclosed. The method utilizes a deep neural network to map the formant structure of whispered audio to a reconstructed harmonic structure. Missing pitch information is inferred from intensity dynamics and semantic context, while speaker identity is maintained through conditioning on a speaker embedding vector. A frame-based generative vocoder processes audio in small segments to allow for …

Automated Service Level Agreement Negotiation And Dynamic Adaptation For Multi-Tenant Submarine Optical Networks, Arvind Kumar, Swarup Kr Ghosh, Richard Moses S, Venkata Narasimhulu Bodagala

Defensive Publications Series

Proposed herein is an automated service level agreement (SLA) negotiation and dynamic adaptation system for multi-tenant optical networks that solves critical problems of service-impacting power fluctuations and operational ambiguity in optical networks. The system provides such features by integrating proactive "make-before-break" algorithms, granular sub-channel monitoring, and novel state management, creating a highly automated, stable, and easy-to-manage shared optical environment.

Universal Power Governance Architecture With Isotopically Enriched Interfaces And Compliant Buffers, N/A

Defensive Publications Series

This disclosure relates to a universal power governance micro-architecture for sub-three nanometer heterogeneous processors. The architecture features an adaptive energy-state engine powered by an isolated sovereign power rail to compute a total energy state of a primary power rail and trigger a transient clamp based on a predictive Hamiltonian function. The system incorporates a semiconductor thermal interface structure with an isotopically enriched diamond layer, which may exceed 99.5% carbon purity, and a refractory metal adhesion layer forming a metal-carbide bond to facilitate ballistic phonon transport. Additionally, the architecture utilizes nanoporous compliant buffer sleeves around conductive vertical pillars to provide reversible …

Cognitive Drift Attacks In Long-Running Ai Agents (How Small Contextual Nudges Quietly Rewrite An Ai System’S Belief System), Pranav Bhatnagar Mr

Defensive Publications Series

Abstract

When AI Does Not Get Hacked, It Slowly Changes Its Mind…

Security discussions around artificial intelligence often assume that attacks must be sudden, visible, and disruptive. Researchers typically search for adversarial prompts, malicious inputs, data poisoning, or explicit manipulation of model outputs. Yet some of the most consequential failures in intelligent systems do not occur through abrupt compromise. Instead, they emerge through gradual shifts in the system’s internal interpretation of reality.

This paper introduces the concept of Cognitive Drift Attacks, a class of adversarial influence in which an AI system’s beliefs, assumptions, or interpretive patterns gradually change over time …

Memory Ghost Attacks In Persistent Ai Systems (When Yesterday’S Context Quietly Hijacks Tomorrow’S Decisions), Pranav Bhatnagar Mr

Defensive Publications Series

Abstract

Why the Most Dangerous AI Attacks May Come From the Past, Not the Present

Artificial intelligence systems are rapidly evolving from simple prompt-response tools into persistent cognitive environments that retain contextual memory across interactions. Modern AI assistants, enterprise copilots, cybersecurity analysis systems, and retrieval-augmented architectures increasingly store fragments of prior conversations, retrieved documents, and contextual reasoning signals to improve decision continuity. While this capability significantly improves usability and system intelligence, it introduces a subtle and largely unexamined security vulnerability residual contextual influence from past interactions.

This paper introduces the concept of Memory Ghost Attacks, a class of adversarial manipulation …

Authority Injection Attacks In Ai Assistants (When Confidence, Titles, And “Expert Sources” Quietly Hijack Ai Reasoning), Pranav Bhatnagar Mr

Defensive Publications Series

Abstract

Why AI Systems Can Be Tricked by Information That Simply Sounds Important

Trust is one of the most powerful forces shaping human decision-making. People frequently evaluate information not only based on evidence, but also based on perceived authority. Titles such as “expert,” “official guideline,” or “research finding” often influence how strongly a claim is believed. Artificial intelligence systems trained on human language patterns inherit many of these same interpretive tendencies. This paper introduces Authority Injection Attacks, a class of adversarial influence techniques in which attackers manipulate AI reasoning by inserting information that appears authoritative or expert-level. Instead of directly …

Zero-Latency Cold Start Via In-Band Application Context Streaming, N/A

Defensive Publications Series

The present disclosure is directed to achieving zero-latency cold starts in cloud computing environments utilizing in-band application context streaming to eliminate disk I/O bottlenecks. When a service becomes idle, a micro-snapshot generating engine performs differential state stripping to identify and serialize only modified memory pages and CPU register values into a micro-snapshot, which is then offloaded to the volatile static random-access memory (SRAM) of a programmable network switch or top-of-rack (ToR) switch. Upon detecting an incoming user request for the hibernated service, the switch may intercept the packet and inject the stored application context as a burst stream of packets …

A System And Method For Real-Time Adaptation And Dynamic Learning Of Artificial Intelligence (Ai) Model, Yu Gu, Sheng Wang, Sam Hamilton

Defensive Publications Series

The present disclosure provides a method and a system for real-time adaptation and dynamic learning of artificial intelligence (AI) model. The method includes incorporating a decoupled multi-model architecture. The multi-model architecture comprises a base module and a feedback module. The method includes training the base module based on the historical transactions data to provide a generalized baseline prediction. The method includes training the feedback module based on the emerging data to provide a real-time prediction. Further, the method includes performing ensemble on the predictions of the base module and the feedback module to provide a reliable and accurate prediction of …

Method, System, And Computer Program Product For Processing A Transaction Using A Location-Based Model, Shawn L. Johnson, Justin Fan, Shizhuo Yu

Defensive Publications Series

Methods, systems, and computer-readable media are disclosed for executing electronic transactions using a location-based predictive model. In one implementation, a location of a user device is monitored and provided to a machine learning model configured to predict a merchant proximate to the device. Merchant information is presented via a user interface, and user input requesting transaction initiation is received. In response, a communication session is automatically established with a merchant application associated with the predicted merchant. The system automatically generates a transaction message comprising user transaction data, credentials, and contextual parameters derived from location and user state, and transmits the …

Optimized Collective Operations For Distributed Machine Learning, N/A

Defensive Publications Series

This publication describes a method for accelerating All-Gather and Reduce-Scatter collective operations in large language models (LLMs) by introducing a new 12-bit floating-point (fp12) datatype, specifically e4m7. This approach aims to reduce network communication overhead by quantizing these operations, thereby decreasing the network payload while maintaining model quality. This document details the problem of communication bandwidth limitations in distributed LLM training and serving, explores related work, and presents experimental results demonstrating the efficacy of the fp12 (e4m7) datatype in various model parallelism configurations.

Orthogonal Memory Projection For Alignment In Stateful Generative Models, Neel Mohbe

Defensive Publications Series

A system and method is needed to mitigate potential alignment erosion in stateful generative models where continuous learning could alter pre-trained safety priors. The technology can introduce a parallel control architecture that may operate alongside a host model to help decouple learning capabilities from alignment constraints. The system can, for example, employ a dual intervention using a runtime residual correction interface to steer immediate output away from undesirable content, and an orthogonal memory projection interface to sanitize signals before they can be written to the model's persistent memory. By mathematically projecting undesirable components out of the memory-write signals, the system …

Contact Center Agentified Topic Analysis, Fergal Keating

Defensive Publications Series

A contact center, sometimes referred to as a call center, is used to help provide solutions for different user inquiries and requests.  For instance, a user (e.g., a customer) can initiate contact or an interaction with a human or artificial intelligence (AI) agent of a contact/call center in which the human/AI agent can help address a possible issue of the user, such as with respect to a product or service involving the user.  This submission proposes specially designed contact center agents, such as a Topic Collector Agent (TCA) and a Topic Routing Agent (TRA) that can operate to gather topics …

Adaptive Peer Hedging For Dynamic Tail Latency Control In Ml Inference Pipelines, Hao Qin

Defensive Publications Series

Method and system for reducing tail latency in machine learning inference pipelines are disclosed. The system includes a primary orchestrator configured to receive an inference request from a client tier component and dispatch the request to a primary execution pool. An adaptive delay controller computes a dynamic hedge delay by optimizing a joint objective balancing tail latency and duplicate execution cost. During the delay interval, a latency risk index (LRI) module computes a LRI using telemetry signals including queue gradients, stage timing deviations, peer load metrics, and network jitter. Upon delay expiry and elevated stall risk, the primary orchestrator forwards …

Methods And Systems For Determining Safe Navigation Routes For A Vessel, Malte Mittendorf, Jesper Dietz, Matthew Nunes, Jacob Hicks, Khushboo Kashyap, Gour Bera, Lauge Kongstad, Guilherme Morgado, Aravind Reddy Pagidela, Martin Merker

Defensive Publications Series

Methods and server systems for determining safe navigation routes for a vessel are disclosed. The method performed by a server system includes receiving route information for operating the vessel. The method includes accessing a plurality of vessel operating parameters recorded at predefined intervals for the vessel. The method includes predicting at least one vessel response parameter for the vessel using at least one Machine Learning (ML) model based on the plurality of vessel operating parameters. Each of the at least one vessel response parameter indicates a predicted response of the vessel within a defined area surrounding the vessel along a …

Cryptographically Binding Authorization Context To Llm And Api Request Execution, Niranjan M M

Defensive Publications Series

Modern API and LLM security solutions focus primarily on detecting malicious requests and deciding whether a request should be allowed or blocked at the API security gateway. While effective for identifying prompt-injection attempts and API abuse, these mechanisms do not provide a reliable way to enforce what an approved request is allowed to do after it leaves the gateway. In real-world LLM systems, requests commonly undergo downstream prompt construction, retrieval-augmented generation (RAG), agent interactions, and tool execution, which may introduce indirect prompt injection and unintended capability escalation beyond the gateway’s visibility.

The proposal introduces a Signed Authorization Context (SAC) mechanism …

Dynamic Memory Management System For Distributed Data Processing Environments, Katyayani Kiranmayee Kolluru, Shivam Mohan, Ravi Shanker Kumar Sinha, Hardik Jain

Defensive Publications Series

The present disclosure relates to a system and method for dynamically managing memory allocation in a distributed data processing platform. The system includes a plurality of workers executing tasks with varying memory requirements, each equipped with a memory sentinel monitoring runtime parameters. A risk engine computes a risk score from telemetry data received from the memory sentinels to anticipate out-of-memory conditions and adjust resource profiles accordingly. When a high risk is detected, the system requests larger resource profiles from a resource manager, provisions new larger workers, and reallocates memory-intensive tasks while preserving their state through a state transfer plane. This …

Trusted And Mutually Authenticated Agentic Ai Communication, Niranjan M M

Defensive Publications Series

Current MCP design carries a wide range of security risks. For example, Industry-leading LLMs may be coerced to use MCP tools and compromise an AI developer’s system through a wide range of attacks viz., Malicious code execution, Remote access control, Credential theft etc.,

Proposed method adds Transaction Database, based on Distributed Hash Table (DHT), where-in every Agentic AI maintains local chain of transactions with other Agentic AI, which are mutually signed. This can be used for auditing of the messages/transactions and telemetry (with the chain of events). As per the proposed method, mutually authenticate the messages (contextual data) between Agentic …

Predicting Permission Prompt Desirability Using On-Device Visual And Textual Content Features, Anh Minh Le, Judith Ulrike Hemp, Enrico Bacis, Balazs Csaba Engedy, Elias Klim, Igor Bilogrevic

Defensive Publications Series

Web browser permission prompts are frequently triggered without sufficient context, leading to user annoyance, decision fatigue, and potential exposure risks. Existing methods for managing these prompts often rely on aggregate site statistics or historical user behavior, which can fail to account for the immediate relevance of a request to the current page content. Such systems may also inadvertently promote deceptive practices if users are frequently misled into granting permissions.

To address these limitations, a method is disclosed for generating a contextual relevance score for permission requests using on-device multimodal analysis. When a permission is requested, a snapshot of the current …

Policy-Bound Governance For Privacy-Preserving Multi-Agent Ai Systems Across Multiple Service Providers, Niranjan M M

Defensive Publications Series

In modern agentic AI environments, autonomous agents dynamically exchange prompts, embeddings, intermediate reasoning states, tool outputs, and memory artifacts, often across geographic regions and infrastructure owned by different providers. Existing solutions govern data primarily at storage boundaries or API entry points, leaving significant gaps during agent reasoning, delegation, and cross-provider execution.

The proposed solution addresses this challenge by binding privacy and governance policies directly to AI artifacts and enforcing those policies at runtime, before any agent action is executed. Each AI artifact carries immutable policy metadata that specifies permitted regions, providers, agent roles, purposes of use, and retention constraints. Autonomous …

Smf Orchestration Engine For Scone Parameter Propagation In 5g Core Networks, Sujin Anagani, Sachin Mosale Vasanthkumar, Charles Jacob, Jis Abraham, Uday Kiran Bokam

Defensive Publications Series

Techniques described herein address a current gap in Fifth Generation (5G) core standards regarding the propagation of Standard Communication with Network Elements (SCONE) parameters between critical network functions by introducing the first standardized mechanism for SCONE parameter exchange in 5G core networks. The techniques introduce a novel Session Management Function (SMF) orchestration framework that intelligently translates policy-level bit-rate directives into SCONE-compliant signal values through dynamic version selection algorithms, logarithmic scaling computations, and capability-aware User Plane Function selection. By defining dedicated Information Elements on the N4 and N7 interfaces and implementing intelligent parameter derivation logic within the SMF, the techniques enable …

Method And System For Adaptive Routing Of Machine Learning Inference Requests Based On Health Scores, Hao Qin

Defensive Publications Series

The present disclosure relates to a method and system for adaptive routing of machine learning inference requests using composite health scores. The method involves maintaining a dynamic health score for each inference engine instance based on multiple performance factors including latency, success rate, timeout occurrence, error rate, saturation, and staleness. Instances are organized in a priority queue according to their health scores, with the instance having the lowest score selected for processing requests. A fairness adjustment is applied to prevent overloading any single instance. Instances that meet specific quarantine criteria are isolated and subject to periodic health assessments, while successfully …

Hypoxia Alert System, Sumukh Ashok Shevde, Abhijit Adsule, Kamran Mustafa

Defensive Publications Series

This document describes a waterproof wearable computing device (e.g., smart watch, smart ring, hearable device, or other body-mounted devices or sensors) that aids in aquatic safety by collecting environmental, biometric, and kinematic data from the user and preemptively alerting the user of a hypoxia induced event (e.g., drowning, unconsciousness, etc.). The wearable computing device uses a multi-sensor fusion approach integrating device components such as barometers, accelerometers, gyroscopes, and physiological sensors to obtain depth data, underwater duration data, kinematic movement data, biometric data, and environmental data, which may be combined with a historical user baseline to determine a lethargy score. The …

Kafka Proactive Gentuner: Intelligent Self-Optimizing Configuration Tuner With Confidence-Scoring, Kartik Goyal, Diwakar Choudary

Defensive Publications Series

[0001] A system and method for managing and optimizing a distributed Apache Kafka production cluster 104. The system 102 collects real-time operational metrics from a plurality of Kafka brokers. The system generates a workload fingerprint and retrieves similar historical fingerprints from a historical case repository. Further, the system identifies top-ranked historical fingerprints using similarity computation. The system generates configuration recommendations if the similarity exceeds a threshold and rejects recommendation generation if the similarity is less than the threshold. Further, the system generates a confidence score for the recommended configuration and compares against a deployment threshold. The system deploys the recommended …

Face-Agnostic Intermittent Loitering Detection In Security Camera Footage, Na

Defensive Publications Series

Security personnel are interested in detecting persons who are loitering in a region captured by a security camera. However, in many scenarios, it is possible that a person walks outside the region captured by a security camera for a period of time before returning to the region. The intermittent presence of the person may not be detected as an instance of loitering. This disclosure describes techniques to detect intermittent loitering in security camera footage. Per the techniques, the period for which a person leaves the screen, if below a certain threshold, is counted towards loitering time. The techniques are face-agnostic, …

Face Agnostic Recognition Of Uniformed Personnel Using Generative Artificial Intelligence, Na

Defensive Publications Series

Footage from security cameras deployed within an organization often includes uniformed employees of the organization, who generally present no security threats. Surfacing the presence of uniformed personnel in security footage can amount to false alarms. This disclosure describes techniques that can identify persons wearing authorized uniforms within security camera footage. Per the techniques, generative artificial intelligence (gen-AI) is used to synthesize images of people wearing various uniforms authorized by the organization, in various poses. Embeddings of the gen-AI images are created and compared with those corresponding to real persons detected in the security camera footage to determine if a detected …

Clustering Person Appearances To Improve Person Recognition, Na

Defensive Publications Series

This disclosure describes techniques that improve the recognition of a person observed in video footage captured by security cameras by leveraging a large number of images of the person obtained from various perspectives from multiple cameras. Images of a person moving through a given space and appearing at various angles and look-directions are obtained and clustered. The cluster of images of a person is subjected to face detection to obtain a gallery of face images for the person. To recognize the person, similarity scores are computed between a reference image of the person and each of the images in the …