Microsoft Technology Licensing, LLC

Idan Yehoshua Hen

Ilay Grossman

Avichai Ben David

Roy Levin

Dotan Patrich

Itay Argoety

Tamer Salman

Andrey Karpovsky

Josef Weizman

Dror Cohen

Moshe Israel

Omer Karin

Naama Kraus

Jonathan Moshe Monsonego

Moshe Shalala

Eran Goldstein

Aharon Naftali Michaels

Rotem Lurie

Avihai Dvir

Amit Magen Medina

Michael Shlomo Navat

Amit Magen

Payal Rani

Sridhar Periyasamy

Ron Keller

Lior Ziv

Amit Dvir Adani

Efrat Reef Guttman

This disclosure relates to utilizing a threat detection system to detect anomalous actions provided by a compromised large generative language model (LLM). For instance, the threat detection system utilizes a detection-based large generative model to process select communication between an application system and the LLM and determine when the LLM may have been potentially compromised. In various implementations, utilizing the detection-based large generative model, the threat detection system determines when an LLM is improperly instructing an application system to invoke tools to perform unapproved actions. Furthermore, when an LLM becomes compromised, the threat detection system intelligently safeguards the detection-based large generative model against similar threats that seek to evade detection or compromise the detection-based large generative model.

Determining anomalous tool invocations by applications using large generative models

Generally discussed herein are devices, systems, and methods for cloud traffic monitoring. A method can include receiving sampled network metadata of a packet transmitted via a computer network, providing the sampled network metadata to a neural network (NN) trained on labeled sampled network metadata, and providing, based on only the sampled network metadata, a classification for the sampled network metadata via the trained neural network.

Detecting network activity from sampled network metadata

Disclosed are techniques for identifying security vulnerabilities in cloud computing resources. To investigate the configuration of cloud resources, a cloud map query is constructed based on a natural language sentence. In some configurations, a machine learning (ML) model converts the sentence to a structured sentence that conveys the intention of the natural language sentence but in machine readable form. The structured sentence is validated to ensure it represents a legal arrangement of cloud resources. The validated structured sentence is then used to generate a cloud map query. The cloud map query retrieves information about a client's cloud resources. In some configurations, query results are analyzed to identify security vulnerabilities and/or follow-up queries are suggested.

Natural language API for security graph exploration

Techniques are described herein that are capable of using weighted peer groups to selectively trigger a security alert. A determination is made that an entity performs an operation. The entity has peers that are categorized among peer groups. For each peer group, an extent to which the peers in the peer group perform the operation is determined. Weights are assigned to the respective peer groups. For each peer group, the extent to which the peers in the peer group perform the operation and the weight that is assigned to the peer group are combined to provide a respective weighted group value. A risk score, which is based at least in part on the weighted group values of the peer groups, is assigned to the operation. The security alert regarding the operation is selectively triggered based at least in part on the risk score.

Using weighted peer groups to selectively trigger a security alert

A computing system is configured for detecting anomalies in deployment configurations of container images at a container network. One or more datasets associated with deployment configurations of a container imager are collected, and a plurality of features are extracted based on the one or more datasets for an ID of the container image. A probability score is then generated based on the plurality of features, using a machine-learning model trained on datasets associated with historical deployment configurations of the container image that have been performed via the container orchestration service. The probability score indicates a probability of whether the deployment configurations of the container image are anomalous or not anomalous when compared historical deployment configurations of the container image. An allow list is generated that includes container images and their respective IDs that have a majority of their deployment configurations that are not anomalous.

Allow list of container images based on deployment configuration at a container orchestration service

A computing system generates from received user input an initial profile. The initial profile specifies expected behavioral patterns of datasets that are to be received by the computing system. The computing system extracts from received datasets features that are indicative of behavioral patterns of the received datasets. The computing system provides the initial profile to first machine-learning models. The first machine-learning models have been trained using a subset of the received datasets. The first machine-learning models use the initial profile to determine if the behavioral patterns of the received datasets are anomalous. The computing system includes second machine-learning models that have been trained using a subset of the received datasets. The second machine-learning models train a second profile based on the extracted features to specify behavioral patterns of the received datasets that are learned by the second machine-learning model.

Machine learning approach for solving the cold start problem in stateful models

A system to detect an abnormal classic authorizations, such as in a classic authorization system of a resource access management system, and take action is described. The system determines an anomaly score in from a model applied to a classic assignment event. An indicator score is determined from the classic assignment event applied to domain-based rules. The security action is taken based on a combination of the anomaly score and the indicator score.

Abnormal classic authorization detection systems

A computing system is configured to detect a request for a deployment of a container at a container orchestration service. One or more datasets associated with the deployment of the container are collected, and a plurality of features associated with the deployment are extracted based on the one or more datasets. A probability score is then generated based on the plurality of features, using a machine-learning model trained on datasets associated with historical deployments of containers that have been performed via the container orchestration service. The probability score indicates a probability that the deployment of the container is anomalous compared to the historical deployments of containers. When the probability score is greater than a threshold, the deployment of the container is determined as anomalous.

Detect anomalous container deployment at a container orchestration service

A system to detect an abnormally permissive role definition, which can include an abnormally permissive custom role definition, and take action is described. The system receives a role definition for a security principal over a scope of resources in which the role definition includes a built-in role and a custom role. Permissions of the role definition and a creation event of the role definition are analyzed. A security score based on the role definition and creation event for the scope of resources is determined. An action is taken based on the security score and the creation event analysis.

Abnormally permissive role definition detection systems

Techniques for user impact potential based security alert management in computer systems are disclosed. One example technique includes receiving an alert indicating that a security rule has been violated by a user. The example technique can also include, in response to receiving the data representing the alert, determining an impact score of the user based on the profile of the user. The impact score represents a deviation of an assigned value to the profile of the user and a mean value of assigned values of profiles of all users in the organization. The example technique can further include calculating a ranking value of the alert in relation to other alerts based on the determine impact score and other impacts scores corresponding to the other alerts and selectively surfacing the alert to a system analyst based on the calculated ranking value in relation to other alerts.

User impact potential for security alert management

According to examples, an apparatus may include a processor and a memory on which is stored machine-readable instructions that when executed by the processor, may cause the processor to determine that an entity was granted an anomalous role assignment to a managed environment. The processor may also, based on the determination that the role assignment of the entity is anomalous, identify at least one indicator associated with the role assignment, determine an indicator value corresponding to the identified at least one indicator, and determine whether the indicator value exceeds a predefined threshold value. The processor may, based on a determination that the indicator value exceeds the predefined threshold value, output an alert indicating that the role assignment is suspicious.

Anomalous and suspicious role assignment determinations

Multiple variate anomaly detection against multiple scopes of the requested resource. Even if one of the variates patterns is sensitive to physical location of the requestor and/or the resource, not all of the variates of the access pattern will be. Furthermore, even if one of the scopes of the resource is sensitive to physical location of the resource, not all scopes will be. Thus, the use of multiple variates of the access pattern and multiple scopes of the anomaly detection allows for better estimates of anomaly detection to be made, even when the source of the access request is virtualized and/or the location of the resource is virtualized.

Multi-variate anomalous access detection

A system to detect abnormal cross authorizations and take action is described. The system determines whether cross authorization event applied to a first trained anomaly detection model and activity post cross authorization event applied to a second trained anomaly detection model is suspicious. An indicator score is determined from rule-based security indications applied to the cross authorization. A security action is taken based on application of the indicator score applied to a threshold.

Abnormal cross authorization detection systems

Technologies are shown for detection of identity misconfiguration that involve collecting identity/role binding and role/access rules data from multiple clusters supported by a computing resource system. Access rules for identities are extracted from the collected data and an access rule prediction model created to predict access rules for identities. An identity definition request for a tenant is received having a requested identity and a role assigned to the identity. A set of access rules is obtained for the role assigned to the identity and a predicted set of access rules is obtained for the requested identity from the prediction model. The access rules for the requested role are compared to the predicted set of access rules and a misconfiguration alert generated when there is a difference between the set of access rules for the requested role and the predicted set of access rules for the requested identity.

Identity misconfiguration detection for role-based access control

A system to determine an intrusion risk and take action is described. The system collaboratively filters a combination based on a user access and a network item in a computer network to determine an associated recommendation score. The system determines connected components of a model of the computer network and separately collaboratively filters the connected components to determine the recommendation score as a measure of intrusion risk. An action is taken on the user access based on the intrusion risk.

Growing community of inventors

Rosh-Ha′ayin, Israel

Idan Yehoshua Hen