Abstract: Zero trust support for secure networks can be provided via a modified virtual private network (VPN) server. For example, the VPN server may receive, from a VPN client executing on a client device, a first access request for a first software application in a computing environment that is accessible via the VPN server. The first access request can include authentication credentials for the VPN server. The VPN server can authenticate the first access request based on the authentication credentials. In response, a first connection tunnel can be provided between the client device and the first software application. The client device can access the first software application via the first connection tunnel. The VPN server can also deny a second access request received via the first connection tunnel for a second software application in the computing environment.

Abstract: Systems and methods are presented to provide a first machine learning model to a collaboration platform. The systems and methods receive a second machine learning model from the collaboration platform that indicates the second machine learning model is based on the first machine learning model. The systems and methods test the second machine learning model using criteria corresponding to the first machine learning model to determine whether the second machine learning model is valid. In turn, the systems and methods publish the second machine learning model to a repository in response to determining that the second machine learning model is valid.

Abstract: A system and method of propagating link aggregation control protocol status to single root input/output virtualization virtual function status. The method includes obtaining, by a processing device and using a relay agent executing on the processing device, state information from a Link Aggregation Control Protocol (LACP) speaker of a host machine. The method includes detecting, based on the state information, a communication failure between a physical function (PF) of a network adapter of a host machine and a first network switch of a plurality of network switches. The method includes identifying one or more computing environments that are communicatively coupled to the first network switch through a first virtual function (VF) associated with the PF. The method includes notifying the one or more computing environments about the communication failure by modifying a link state of the first VF.

Abstract: This disclosure provides devices, systems, methods, and techniques for booting up an operating system with an initial scalable filesystem. An example method includes loading, from a non-transitory and non-volatile memory, a bootloader into a random access memory (RAM). The method includes loading, by the bootloader, a kernel into the RAM. The method includes initializing, by the kernel, an initial filesystem to load storage drivers relevant to the hardware resources for starting essential storage components. The method includes, when the initial filesystem is initialized, mounting, by a processing device, an overlay filesystem onto the initial filesystem of the RAM. The method further includes transitioning, upon completion of system initialization tasks in the overlay filesystem, from the overlay filesystem to a target root filesystem.

Abstract: The size of an input prompt for a large language model (LLM) can be reduced according to some techniques described herein. As one example, a system can generate a first vector embedding of a natural language query from a user. The system can also generate a second vector embedding of a conversation history between the user and a LLM. The system can then generate a truncated embedding by executing a filtering process based on the first vector embedding and the second vector embedding. The system can generate an input prompt based on the truncated embedding. The system can then execute the LLM based on the input prompt to generate a response to the query. Reducing the size of the input prompt can decrease latency, execution time, and hallucinations associated with the LLM.

Abstract: Containers can be migrated from a cloud environment to a local system. For example, a system can receive a first version of a container stored in a first cloud environment and a second version of the container stored in a second cloud environment. The first version may include a first set of container layers and the second version may include a second set of container layers. The system may determine a container layer of the first set of container layers that does not match a corresponding container layer of the second set of container layers. The system may then select the container layer or the corresponding container layer. Additionally, the system may generate a third set of container layers that includes the selected container layer and that includes container layers of the first set of container layers that match corresponding container layers in the second set of container layers.

Abstract: Techniques that provide source metadata associated with code recommendations generated by an automation controller are described. A training data set comprising queries, code recommendations and source metadata associated with each of the code recommendations may be stripped of the source metadata and encoded by an NLP model to generate vector representations of each code recommendation. Each vector representation is stored in an index in association with the source metadata corresponding to the underlying code recommendation. In response to receiving a query, the query is processed using an inference model to generate a second code recommendation, which is encoded using the NLP model to generate a second vector representing the second code recommendation. The second vector is compared to the index to identify the source metadata corresponding to the second code recommendation. The second code recommendation and corresponding source metadata are provided via an interface of the automation controller.

Abstract: Techniques for executing a software program that uses a different memory-page size than its host computing environment are described. As one example, a system can deploy a virtual machine in a host computing environment of the computing device. The host computing environment can include a host operating system that is configured to use a first memory page size. The virtual machine can be configured to use a second memory page size that is different from the first memory page size. Based on a software program being configured to use the second memory page size, the system can execute the software program in the virtual machine.

Abstract: A system can be provided for modifying cluster configuration settings using machine learning. For example, the system can determine a numerical value representative of a configuration setting at an active computing cluster. The system can further compute a set of similarity scores using the numerical value. Each similarity score in the set of similar scores can be indicative of a level of similarity of the active computing cluster to each of a set of computing clusters with respect to the configuration setting. The system can further select, based on the set of similarity scores and using a machine learning model, a subset of computing clusters from the set of computing clusters. The system can then generate a recommended modification to the configuration setting based on the subset of computing clusters. Additionally, the system can execute a modification operation to implement the recommended modification to the configuration setting.

Abstract: Training nodes can be selected for use in training a machine-learning model according to some aspects described herein. In one example, a system can receive performance-metric values generated by training nodes, where the training nodes are configured to generate the performance-metric values by implementing an evaluation phase in which the training nodes partially train models using first training data. The system can select a subset of the training nodes based on the performance-metric values. The system can then transmit commands to the subset of training nodes for causing the subset of training nodes to implement a training phase in which the subset of training nodes further train the models using second training data.

Abstract: A method includes receiving, by a processing device, a first packet addressed to a first virtualized execution environment, determining, by the processing device, whether the first packet has similar characteristics with a second packet by applying a first filtering rule to the first packet, wherein the first filtering rule is generated in view of characteristics of the second packet, and wherein the second packet is stored in a first filtering queue of a second virtualized execution environment, and responsive to determining that the first packet is similar to the second packet, discarding, by the processing device, the first packet.

Abstract: A system can be provided for managing software service operations in multi-node environments (e.g., edge computing systems). For example, the system may receive a request to execute an external event from an input device of an edge computing system. In response to receiving the request to execute the external event, the system may identify one or more operations involved in execution of the external event based on configuration data associated with the edge computing system. The system may then control one or more computing nodes to execute the external event by causing the computing nodes to perform the operations.

Abstract: This application relates to automatic real-time detection and mitigation of reverse shell attacks. In some examples, a system can execute a machine-learning model to detect a reverse shell attempt by an attacker in relation to a target network machine in a computer network. In response to detecting the reverse shell attempt, the system can deploy a honeypot network machine in the computer network. The honeypot network machine can be configured to mimic the target network machine and can be sandboxed from a remainder of the computer network. The system can then redirect a connection of the attacker from the target network machine to the honeypot network machine. By redirecting the attacker to the sandboxed honeypot machine, the reverse shell attack can be mitigated.

Abstract: In one example, a system can receive a dependency tree associated with a first software program. The dependency tree can indicate that the first software program has a first dependency on a second software program and a second dependency on a third software program. The dependency tree can further indicate that the third software program has a third dependency on the second software program. The system can determine an update to the third dependency for the third software program. The update can be from relying on the second software program to relying on a fourth software program. The system can cause an action associated with the first software program based on determining the update.

Abstract: Freeze attacks can be prevented using the techniques described herein. In one example, a computing device can receive a software update for a software program. The computing device can receive the software update from an update provider through a first communication channel. Based on receiving the software update, the computing device can transmit a request associated with the software update to a remote server through a second communication channel, the second communication channel being different from the first communication channel. The computing device can receive a response to the request from the remote server. The response can indicate that the software update is outdated. Based on the response indicating that that software update is outdated, the computing device can discard the software update from the computing device without installing the software update. Discarding the software update in this way can help prevent against a freeze attack.

Abstract: A retrieval augmented generation (RAG) process for a large language model (LLM) can be enhanced using contextual data. For example, a system can transform a query into a query embedding and generate weighted similarity scores for chunk embeddings stored in a database. Each weighted similarity score can be generated based on the query embedding, the chunk embedding, and at least one contextual data embedding. The system can select at least one of the chunk embeddings based on the weighted similarity scores. The at least one chunk embedding corresponds to at least one chunk of at least one text document. The system can then retrieve the at least one text document from a document database, generate an input for a LLM that includes the at least one text document and the query, and execute the LLM based on the input to generate a response to the query.

Abstract: Systems and methods for detecting an interaction with an application displayed on a user interface to perform a metric exporting function of the application by using an exporter configuration that is inaccessible to a user. The system and method also includes exporting by a processing device executing an exporter, metrics generated by each of the one or more classes of the application to a monitoring program, wherein the exporter uses the exporter configuration to export the metrics generated by each of the one or more classes.

Abstract: Service disruption avoidance techniques are provided based on deployment metadata associated with a cluster. The cluster includes one or more nodes. A plurality of applications are installed on the cluster. Prior to a maintenance activity to the one or more nodes of the cluster, a processor receives a request to determine a state of maintenance availability of an application. The processor retrieves deployment metadata for the plurality of applications installed on the cluster, in response to receiving the request. The processor parses the deployment metadata to identify one or more installation rules associated with the plurality of applications. The processor determines a set of conflicts for the first application, and determines the state of maintenance availability of the first application based on the set of conflicts.

Abstract: The size of an input prompt for a large language model (LLM) can be reduced according to some techniques described herein. As one example, a system can generate a first vector embedding of a natural language query from a user. The system can also generate a second vector embedding of a conversation history between the user and a LLM. The system can then generate a truncated embedding by executing a filtering process based on the first vector embedding and the second vector embedding. The system can generate an input prompt based on the truncated embedding. The system can then execute the LLM based on the input prompt to generate a response to the query. Reducing the size of the input prompt can decrease latency, execution time, and hallucinations associated with the LLM.

Abstract: Replicated secrets can be migrated from a cloud environment to a local system. For example, a system can receive a secret stored in a first geographic region and a second geographic region of a cloud environment. The system can determine a first security level associated with a first source location of the secret and a second security level associated with a second source location of the secret. The system can encrypt the secret using a first encryption key associated with the first security level to generate a first encrypted secret. The system can determine that a first filesystem in a physical server corresponds to the first source location of the secret and store the first encrypted secret in the first filesystem of the physical server. The first encrypted secret can be usable to control access of one or more protected computing resources.