F5 Distributed Cloud – CE High Availability Options: A Comparative Exploration
This article explores an alternative approach to achieve HA across single CE nodes, catering for use cases requiring higher performance and granular control over redundancy and failover management. Introduction F5 Distributed Cloud offers different techniques to achieve High Availability (HA) for Customer Edge (CE) nodes in an active-active configuration to provide redundancy, scaling on-demand and simplify management. By default, F5 Distributed Cloud uses a method for clustering CE nodes, in which CEs keep track of peers by sending heartbeats and facilitating traffic exchange among themselves. This method also handles the automatic transfer of traffic, virtual IPs, and services between CE peers —excellent for simplified deployment and running App Stack sites hosting Kubernetes workloads. However, if CE nodes are deployed mainly to manage L3/L7 traffic and application security, this default model might lack the flexibility needed for certain scenarios. Many of our customers tell us that achieving high availability is not so straightforward with the current clustering model. These customers often have a lot of experience in managing redundancy and high availability across traditional network devices. They like to manage everything themselves—from scheduling when to switch over to a redundant pair (planned failover), to choosing how many network paths (tunnels) to use between CEs to REs (Regional Edges) or other CEs. They also want to handle any issues device by device, decide the number of CE nodes in a redundancy group, and be able to direct traffic to different CEs when one is being updated. Their feedback inspired us to write this article, where we explore a different approach to achieve high availability across CEs. The default clustering model is explained in this document: https://docs.cloud.f5.com/docs/ves-concepts/site#cluster-of-nodes Throughout this article, we will dive into several key areas: An overview of the default CE clustering model, highlighting its inherent challenges and advantages. Introduction to an alternative clustering strategy: Single Node Clustering, including: An analysis of its challenges and benefits. Identification of scenarios where this approach is most applicable. A guide to the configuration steps necessary to implement this model. An exploration of failover behavior within this framework. A comparison table showing how this new method differs from the default clustering method. By the end of this article, readers will gain an understanding of both clustering approaches, enabling informed decisions on the optimal strategy for their specific needs. Default CE Clustering Overview In a standard CE clustering setup, a cluster must have at least three Master nodes, with subsequent additions acting as Worker nodes. A CE cluster is configured as a "Site," centralizing operations like pool configuration and software upgrades to simplify management. In this clustering method, frequent communication is required between control plane components of the nodes on a low latency network. When a failover happens, the VIPs and services - including customer’s compute workloads - will transition to the other active nodes. As shown in the picture above, a CE cluster is treated as a single site, regardless of the number of nodes it contains. In a Mesh Group scenario, each mesh link is associated with one single tunnel connected to the cluster. These tunnels are distributed among the master nodes in the cluster, optimizing the total number of tunnels required for a large-scale Mesh Group. It also means that the site will be connected to REs only via 2 tunnels – one to each RE. Design Considerations for Default CE Clustering model: Best suited for: 1- App Stack Sites: Running Kubernetes workloads necessitates the default clustering method for container orchestration across nodes. 2- Large-scale Site-Mesh Groups (SMG) 3- Cluster-wide upgrade preference: Customers who favour managing nodes collectively will find cluster-wide upgrades more convenient, however without control over the upgrade sequence of individual nodes. Challenges: o Network Bottleneck for Ingress Traffic: A cluster connected to two Regional Edge (RE) sites via only 2 tunnels can lead to only two nodes processing external (ingress) traffic, limiting the use of additional nodes to process internal traffic only. o Three-master node requirement: Some customers are accustomed to dual-node HA models and may find the requirement for three master nodes resource-intensive. o Hitless upgrades: Controlled, phased upgrades are preferred by some customers for testing before widespread deployment, which is challenging with cluster-wide upgrades. o Cross-site deployments: High network latency between remote data centers can impact cluster performance due to the latency sensitivity of etcd daemon, the backbone of cluster state management. If the network connection across the nodes gets disconnected, all nodes will most likely stop the operation due to the quorum requirements of etcd. Therefore, F5 recommends deploying separate clusters for different physical sites. o Service Fault Sprawl and limited Node fault tolerance: Default clusters can sometimes experience a cascading effect where a fault in a node spreads throughout the cluster. Additionally, a standard 3-node cluster can generally only tolerate the failure of two nodes. If the cluster was originally configured with three nodes, functionality may be lost if reduced to a single active node. These limitations stem from the underlying clustering design and its dependency on etcd for maintaining cluster state. The Alternative Solution: HA Between Multiple Single Nodes The good news is that we can achieve the key objectives of the clustering – which are streamlined management and high availability - without the dependency on the control plane clustering mechanisms. Streamlined management using “Virtual Site”: F5 Distributed Cloud provides a mechanism called “Virtual Site” to perform operations on a group of sites (site = node or cluster of nodes), reducing the need to repeat the same set of operations for each site. The “Virtual Site” acts as an abstraction layer, grouping nodes tagged with a unique label and allows collectively addressing these nodes as a single entity. Configuration of origin pools and load balancers can reference Virtual Sites instead of individual sites/nodes, to facilitate cluster-like management for two or more nodes and enabling controlled day 2 operations. When a node is disassociated from Virtual Site by removing the label, it's no longer eligible for new connections, and its listeners are simultaneously deactivated. Upgrading nodes is streamlined: simply remove the node's label to exclude it from the Virtual Site, perform the upgrade, and then reapply the label once the node is operational again. This procedure offers you a controlled failover process, ensuring minimal disruption and enhanced manageability by minimizing the blast radius and limiting the cope of downtime. As traffic is rerouted to other CEs, if something goes wrong with an upgrade of a CE node, the services will not be impacted. HA/Redundancy across multiple nodes: Each single node in a Virtual Site connect to dual REs through IPSec or SSL/TLS tunnels, ensuring even load distribution and true active-active redundancy. External (Ingress) Traffic: In the Virtual Site model, the Regional Edges (REs) distribute external traffic evenly across all nodes. This contrasts with the default clustering approach where only two CE nodes are actively connected to the REs. The main Virtual Site advantage lies in its true active/active configuration for CEs, increasing the total ingress traffic capacity. If a node becomes unavailable, the REs will automatically reroute the new connections to another operational node within the Virtual Site, and the services (connection to origin pools) remain uninterrupted. Internal (East-West) Traffic: For managing internal traffic within a single CE node in a Virtual Site (for example, when LB objects are configured to be advertised within the local site), all network techniques applicable to the default clustering model can be employed in this model as well, except for the Layer 2 attachment (VRRP) method. Preferred load distribution method for internal traffic across CEs: Our preferred methods for load balancing across CE nodes are either DNS based load balancing or Equal-Cost Multi-Path (ECMP) routing utilizing BGP for redundancy. DNS Load Balancer Behavior: If a node is detached from a Virtual Site, its associated listeners and Virtual IPs (VIPs) are automatically withdrawn. Consequently, the DNS load balancer's health checks will mark those VIPs as down and prevent them from receiving internal network traffic. Current limitation for custom VIP and BGP: When using BGP, please note a current limitation that prevents configuring a custom VIP address on the Virtual Site. As a workaround, custom VIPs should be advertised on individual sites instead. The F5 product team is actively working to address this gap. For a detailed exploration of traffic routing options to CEs, please refer to the following article here: https://community.f5.com/kb/technicalarticles/f5-distributed-cloud---customer-edge-site---deployment--routing-options/319435 Design Considerations for Single Node HA Model: Best suited for: 1- Customers with high throughput requirement: This clustering model ensures that all Customer Edge (CE) nodes are engaged in managing ingress traffic from Regional Edges (REs), which allows for scalable expansion by adding additional CEs as required. In contrast, the default clustering model limits ingress traffic processing to only two CE nodes per cluster, and more precisely, to a single node from each RE, regardless of the number of worker nodes in the cluster. Consequently, this model is more advantageous for customers who have high throughput demands. 2- Customers who prefer to use controlled failover and software upgrades This clustering model enables a sequential upgrade process, where nodes are updated individually to ensure each node upgrades successfully before moving on to the other nodes. The process involves detaching the node from the cluster by removing its site label, which causes redirecting traffic to the remaining nodes during the upgrade. Once upgraded, the label is reapplied, and this process is repeated for each node in turn. This is a model that customers have known for 20+ years for upgrade procedures, with a little wrinkle with the label. 3- Customers who prefer to distribute the load across remote sites Nodes are deployed independently and do not require inter-node heartbeat communication, unlike the default clustering method. This independence allows for their deployment across various data centers and availability zones while being managed as a single entity. They are compatible with both Layer 2 (L2) spanned and Layer 3 (L3) spanned data centers, where nodes in different L3 networks utilize distinct gateways. As long as the nodes can access the origin pools, they can be integrated into the same "Virtual Site". This flexibility caters to customers' traditional preferences, such as deploying two CE nodes per location, which is fully supported by this clustering model. Challenges: Lack of VRRP Support: The primary limitation of this clustering method is the absence of VRRP support for internal VIPs. However, there are some alternative methods to distribute internal traffic across CE nodes. These include DNS based routing, BGP with Equal-Cost Multi-Path (ECMP) routing, or the implementation of CEs behind another Layer 4 (L4) load balancer capable of traffic distribution without source address alteration, such as F5 BIG-IPs or the standard load balancers provided by Azure or AWS. Limitation on Custom VIP IP Support: Currently, the F5 Distributed Cloud Console has a restriction preventing the configuration of custom virtual IPs for load balancer advertisements on Virtual Sites. We anticipate this limitation will be addressed in future updates to the F5 Distributed Cloud platform. As a temporary solution, you can advertise the LB across multiple individual sites within the Virtual Site. This approach enables the configuration of custom VIPs on those sites. Requires extra steps for upgrading nodes Unlike the Default clustering model where upgrades can be performed collectively on a group of nodes, this clustering model requires upgrading nodes on an individual basis. This may introduce more steps, especially in larger clusters, but it remains significantly simpler than traditional network device upgrades. Large-Scale Mesh Group: In F5 Distributed Cloud, the "Mesh Group" feature allows for direct connections between sites (whether individual CE sites or clusters of CEs) and other selected sites through IPSec tunnels. For CE clusters, tunnels are established on a per-cluster basis. However, for single-node sites, each node creates its own tunnels to connect with remote CEs. This setup can lead to an increased number of tunnels needed to establish the mesh. For example, in a network of 10 sites configured with dual-CE Virtual Sites, each CE is required to establish 18 IPSec tunnels to connect with other sites, or 19 for a full mesh configuration. Comparatively, a 10-site network using the default clustering method—with a minimum of 3 CEs per site—would only need up to 9 tunnels from each CE for full connectivity. Opting for Virtual Sites with dual CEs, a common choice, effectively doubles the number of required tunnels from each CE when compared to the default clustering setup. However, despite this increase in tunnels, opting for a Mesh configuration with single-node clusters can offer advantages in terms of performance and load distribution. Note: Use DC Groups as an alternative solution to Secure Mesh Group for CE connectivity: For customers with existing private connectivity between their CE nodes, running Site Mesh Group (SMG) with numerous IPsec tunnels can be less optimal. As a more scalable alternative for these customers, we recommend using DC Cluster Group (DCG). This method utilizes IP-in-IP tunnels over the existing private network, eliminating the need for individual encrypted IPsec tunnels between each node and streamlining communication between CE nodes via IP-n-IP encapsulations. Configuration Steps The configuration for creating single node clusters involves the following steps: Creating a Label Creating a Virtual Site Applying the label to the CE nodes (sites) Review and validate the configuration The detailed configuration guide for the above steps can be found here: https://docs.cloud.f5.com/docs/how-to/fleets-vsites/create-virtual-site Example Configuration: In this example, you can create a label called "my-vsite" to group CE nodes that belong to the same Virtual Site. Within this label, you can then define different values to represent different environments or clusters, such as specific Azure region or an on-premise data center. Then a Virtual Site of “CE” type can be created to represent the CE cluster in “Azure-AustraliaEast-vSite" and tied to any CE that is tagged with the label “my-vsite=Azure-AustraliaEast-vSite”: Now, any CE node that should join the cluster (Virtual Site), should get this label: Verification: To confirm the Virtual Site configuration is functioning as intended, we joined two CEs (k1-azure-ce2 and k1-azure-ce03) into the Virtual Site and evaluated the routing and load balancing behavior. Test 1: Public Load Balancer (Virtual Site referenced in the pool) The diagram shows a public "Load Balancer" advertised on the RE referencing a pool that uses the newly created Virtual Site to access the private application: As shown below, the pool member was configured to be accessed through the Virtual Site: Analysis of the request logs in the Performance dashboard confirmed that all requests to the public website were evenly distributed across both CEs. Test 2: Internal Load Balancer (LB advertised on the Virtual Site) We deployed an internal Load Balancer and advertised it on the newly created Virtual Site, utilizing the pool that also references the same Virtual Site (k1-azure-ce2 and k1-azure-ce03). As shown below, the Load Balancer was configured to be advertised on the Virtual Site. Note: Here we couldn't use a "shared" custom VIP across the Virtual Site due to a current platform constraint. If a custom VIP is required, we should use "site" as opposed to "Virtual Site" and advertise the Load Balancer on all sites, like below picture: Request logs revealed that when traffic reached either CE node within the Virtual Site, the request was processed and forwarded locally to the pool member. In the example below: src_site: Indicates the CE (k1-azure-ce2) that processed the request. src_ip: Represents the client's source IP address (192.168.1.68). dst_site: Indicates the CE (k1-azure-ce2) from which the pool member is accessed. dst_ip: Represents the IP address of the pool member (192.168.1.6). Resilience Testing: To assess the Virtual Site's resilience, we intentionally blocked network access from k1-azure-ce2 CE to the pool member (192.168.1.6). The CE automatically rerouted traffic to the pool member via the other CE (k1-azure-ce03) in the Virtual Site. Note:By default, CEs can communicate with each other via the F5 Global Network. This can be customized to use direct connectivity through tunnels if the CEs are members of the same DC Cluster Group (IP-n-IP tunneling) or Secure Mesh Groups (IPSec tunneling). The following picture shows the traffic flow via F5 Global Network. The following picture shows the traffic flow via the IP-n-IP tunnel when a DC Clustering Group (DCG) is configured across the CE nodes. Failover Behaviour When a CE node is tied to a Virtual Site, all internal Load Balancers (VIPs) advertised on that Virtual Site will be deployed in the CE. Additionally, the Regional Edge (RE) begins to use this node as one of the potential next hops for connections to the origin pool. Should the CE become unavailable, or if it lacks the necessary network access to the origin server, the RE will almost seamlessly reroute connections through the other operational CEs in the Virtual Site. Uncontrolled Failover: During instances of uncontrolled failover, such as when a node is unexpectedly shut down from the hypervisor, we have observed a handful of new connections experiencing timeouts. However, these issues were resolved by implementing health checks within the origin pool, which prevented any subsequent connection drops. Note: Irrespective of the clustering model in use, it's always recommended to configure health checks for the origin pool. This practice enhances failover responsiveness and mitigates any additional latency incurred during traffic rerouting. Controlled Failover: The moment a CE node is disassociated from the Virtual Site — by the removal of its label— the CE node will not be used by RE to connect to origin pools anymore. At the same time, all Load Balancer listeners associated with that Virtual Site are withdrawn from the node. This effectively halts traffic processing for those applications, preventing the node from receiving related traffic. During controlled failover scenarios, we have observed seamless service continuity on externally advertised services (to REs). On-Demand Scaling: F5 Distributed Cloud provides a flexible solution that enables customers to scale the number of active CE nodes according to demand. This allows you to easily add more powerful CE nodes during peak periods (such as promotional events) and then remove them when demand subsides. With the Virtual Sites method, you can even mix and match node sizes within your cluster (Virtual Site), providing granular control over resources. It's advisable to monitor CE node performance and implement node related alerts. These alerts notify you when nodes are operating at high capacity, allowing for timely addition of extra nodes as needed. Moreover, you can monitor node’s health in the dashboard. CPU, Memory and Disk utilizations of nodes can be a good factor in determining if more nodes are needed or not. Furthermore, the use of Virtual Sites makes managing this process even easier, thanks to labels. Node Based Alerts: Node-based alerts are essential for maintaining efficient CE operations. Accessing the alerts in the Console: To view alerts, go to Multi-Cloud Network Connect > Notifications > Alerts. Here, you can see both "Active Alerts" and "All Alerts." Alerts related to node health fall under the "infrastructure" alert group. The following screenshot shows alerts indicating high loads on the nodes. Configuring Alert Policies: Alert policies determine the notification process for raised alerts. To set up an alert policy, navigate to Multi-Cloud Network Connect > Alerts Management > Alert Policies. An alert policy consists of two main elements: the alert receiver configuration and the policy rules. Configuring Alert Receiver: The configuration allows for integration with platforms like Slack and PagerDuty, among others, facilitating notifications through commonly used channels. Configuring Alert Rules: For alert selection, we recommend configuring notifications for alerts with severity of “Major” or “Critical” at a minimum. Alternatively, the “infrastructure” group which includes node-based alerts can be selected. Comparison Table Criteria Default Cluster Single Node HA Minimum number of nodes in HA 3 2 Upgrade operations Per cluster Per Node Network redundancy and client side routing for east-west traffic VRRP, BGP, DNS, L4/7 LB DNS, L4/7 LB, BGP* Tunnels to RE 2 tunnels per cluster 2 tunnels per node Tunnels to other CEs (SMG or DCG) 1 tunnel from each cluster 1 tunnel from each node External traffic processing Limited to 2 nodes All nodes will be active Internal traffic processing All nodes can be active All nodes can be active Scale management in Public Cloud Sites Straightforward, by configuring ingress interfaces in Azure/AWS/GCP sites Straightforward, by adding or removing the labels Scale management in Secure Mesh Sites Requires reconfiguring the cluster (secure mesh site) - may cause interruption Straightforward, by adding or removing the labels Custom VIP IP Available Not Available (Planned to be available in future releases), workaround available. Node sizes All nodes should be same size. Upgrading node size in a cluster is a disruptive operation. Any node sizes or clusters can join the Virtual Site * When using BGP, please note a current limitation that prevents configuring custom VIP address on the Virtual Site. Conclusion: F5 Distributed Cloud offers a flexible approach to High Availability (HA) across CE nodes, allowing customers to select the redundancy model that best fits their specific use cases and requirements. While we continue to advocate for default clustering approach due to their operational simplicity and shared VRRP VIP or, unified network configuration benefits, especially for routine tasks like upgrades, the Virtual Site and single node HA model presents some great use cases. It not only addresses the limitations and challenges of the default clustering model, but also introduces a solution that is both scalable and adaptable. While Virtual Sites offer their own benefits, we recognize they also present trade-offs. The overall benefits, particularly for scenarios demanding high ingress (RE to CE) throughput and controlled failover capabilities cater to specific customer demands. The F5 product and development team remains committed to addressing the limitations of both default clustering and Virtual Sites discussed throughout this article. Their focus is on continuous improvement and finding the solutions that best serve our customers' needs. References and Additional Links: Default Clustering model: https://docs.cloud.f5.com/docs/ves-concepts/site#cluster-of-nodes Configuration guide for Virtual Sites:https://docs.cloud.f5.com/docs/how-to/fleets-vsites/create-virtual-site Routing Options for CEs:https://community.f5.com/kb/technicalarticles/f5-distributed-cloud---customer-edge-site---deployment--routing-options/319435 Configuration guide for DC Clustering Group:https://docs.cloud.f5.com/docs/how-to/advanced-networking/configure-dc-cluster-groupSecuring the LLM User Experience with an AI Firewall
As artificial intelligence (AI) seeps into the core day-to-day operations of enterprises, a need exists to exert control over the intersection point of AI-infused applications and the actual large language models (LLMs) that answer the generated prompts. This control point should serve to impose security rules to automatically prevent issues such as personally identifiable information (PII) inadvertently exposed to LLMs. The solution must also counteract motivated, intentional misuse such as jailbreak attempts, where the LLM can be manipulated to provide often ridiculous answers with the ensuing screenshotting attempting to discredit the service. Beyond the security aspect and the overwhelming concern of regulated industries, other drivers include basic fiscal prudence 101, ensuring the token consumption of each offered LLM model is not out of hand. This entire discussion around observability and policy enforcement for LLM consumption has given rise to a class of solutions most frequently referred to as AI Firewalls or AI Gateways (AI GW). An AI FW might be leveraged by a browser plugin, or perhaps applying a software development kit (SDK) during the coding process for AI applications. Arguably, the most scalable and most easily deployed approach to inserting AI FW functionality into live traffic to LLMs is to use a reverse proxy. A modern approach includes the F5 Distributed Cloud service, coupled with an AI FW/GW service, cloud-based or self-hosted, that can inspect traffic intended for LLMs like those of OpenAI, Azure OpenAI, or privately operated LLMs like those downloaded from Hugging Face. A key value offered by this topology, a reverse proxy handing off LLM traffic to an AI FW, which in turn can allow traffic to reach target LLMs, stems from the fact that traffic is seen, and thus controllable, in both directions. Should an issue be present in a user’s submitted prompt, also known as an “inference”, it can be flagged: PII (Personally Identifiable Information) leakage is a frequent concern at this point. In addition, any LLM responses to prompts are also seen in the reverse path: consider a corrupted LLM providing toxicity in its generated replies. Not good. To achieve a highly performant reverse proxy approach to secured LLM access, a solution that can span a global set of users, F5 worked with Prompt Security to deploy an end-to-end AI security layer. This article will explore the efficacy and performance of the live solution. Impose LLM Guardrails with the AI Firewall and Distributed Cloud An AI firewall such as the Prompt Security offering can get in-line with AI LLM flows through multiple means. API calls from Curl or Postman can be modified to transmit to Prompt Security when trying to reach targets such as OpenAI or Azure OpenAI Service. Simple firewall rules can prevent employee direct access to these well-known API endpoints, thus making the Prompt Security route the sanctioned method of engaging with LLMs. A number of other methods could be considered but have concerns. Browser plug-ins have the advantage of working outside the encryption of the TLS layer, in a manner similar to how users can use a browser’s developer tools to clearly see targets and HTTP headers of HTTPS transactions encrypted on the wire. Prompt Security supports plugins. A downside, however, of browser plug-ins is the manageability issue, how to enforce and maintain across-the-board usage, simply consider the headache non-corporate assets used in the work environment. Another approach, interesting for non-browser, thick applications on desktops, think of an IDE like VSCode, might be an agent approach, whereby outbound traffic is handled by an on-board local proxy. Again, Prompt can fit in this model however the complexity of enforcement of the agent, like the browser approach, may not always be easy and aligned with complete A-to-Z security of all endpoints. One of the simplest approaches is to ingest LLM traffic through a network-centric approach. An F5 Distributed Cloud HTTPS load balancer, for instance, can ingest LLM-bound traffic, and thoroughly secure the traffic at the API layer, things like WAF policy and DDoS mitigations, as examples. HTTP-based control plane security is the focus here, as opposed to the encapsulated requests a user is sending to an LLM. The HTTPS load balancer can in turn hand off traffic intended for the likes of OpenAI to the AI gateway for prompt-aware inspections. F5 Distributed Cloud (XC) is a good architectural fit for inserting a third-party AI firewall service in-line with an organization’s inferencing requests. Simply project a FQDN for the consumption of AI services; in this article we used the domain name “llmsec.busdevF5.net” into the global DNS, advertising one single IP address mapping to the name. This DNS advertisement can be done with XC. The IP address, through BGP-4 support for anycast, will direct any traffic to this address to the closest of 27 international points of presence of the XC global fabric. Traffic from a user in Asia may be attracted to Singapore or Mumbai F5 sites, whereas a user in Western Europe might enter the F5 network in Paris or Frankfurt. As depicted, a distributed HTTPS load balancer can be configured – “distributed” reflects the fact traffic ingressing in any of the global sites can be intercepted by the load balancer. Normally, the server name indicator (SNI) value in the TLS Client Hello can be easily used to pick the correct load balancer to process this traffic. The first step in AI security is traditional reverse proxy core security features, all imposed by the XC load balancer. These features, to name just a few, might include geo-IP service policies to preclude traffic from regions, automatic malicious user detection, and API rate limiting; there are many capabilities bundled together. Clean traffic can then be selected for forwarding to an origin pool member, which is the standard operation of any load balancer. In this case, the Prompt Security service is the exclusive member of our origin pool. For this article, it is a cloud instantiated service - options exist to forward to Prompt implemented on a Kubernetes cluster or running on a Distributed Cloud AppStack Customer Edge (CE) node. Block Sensitive Data with Prompt Security In-Line AI inferences, upon reaching Prompt’s security service, are subjected to a wide breadth of security inspections. Some of the more important categories would include: Sensitive data leakage, although potentially contained in LLM responses, intuitively the larger proportion of risk is within the requesting prompt, with user perhaps inadvertently disclosing data which should not reach an LLM Source code fragments within submissions to LLMs, various programming languages may be scanned for and blocked, and the code may be enterprise intellectual property OWASP LLM top 10 high risk violations, such as LLM jailbreaking where the intent is to make the LLM behave and generate content that is not aligned with the service intentions; the goal may be embarrassing “screenshots”, such as having a chatbot for automobile vendor A actually recommend a vehicle from vendor B OWASP Prompt Injection detection, considered one of the most dangerous threats as the intention is for rogue users to exfiltrate valuable data from sources the LLM may have privileged access to, such as backend databases Token layer attacks, such as unauthorized and excessive use of tokens for LLM tasks, the so-called “Denial of Wallet” threat Content moderation, ensuring a safe interaction with LLMs devoid of toxicity, racial and gender discriminatory language and overall curated AI experience aligned with those productivity gains that LLMs promise To demonstrate sensitive data leakage protection, a Prompt Security policy was active which blocked LLM requests with, among many PII fields, a mailing address exposed. To reach OpenAI GPT3.5-Turbo, one of the most popular and cost-effective models in the OpenAI model lineup, prompts were sent to an F5 XC HTTPS load balancer at address llmsec.busdevf5.net. Traffic not violating the comprehensive F5 WAF security rules were proxied to the Prompt Security SaaS offering. The prompt below clearly involves a mailing address in the data portion. The ensuing prompt is intercepted by both the F5 and Prompt Security solutions. The first interception, the distributed HTTPS load balancer offered by F5 offers rich details on the transaction, and since no WAF rules or other security policies are violated, the transaction is forwarded to Prompt Security. The following demonstrates some of the interesting details surrounding the transaction, when completed (double-click to enlarge). As highlighted, the transaction was successful at the HTTP layer, producing a 200 Okay outcome. The traffic originated in the municipality of Ashton, in Canada, and was received into Distributed Cloud in F5’s Toronto (tr2-tor) RE site. The full details around the targeted URL path, such as the OpenAI /v1/chat/completions target and the user-agent involved, vscode-restclient, are both provided. Although the HTTP transaction was successful, the actual AI prompt was rejected, as hoped for, by Prompt Security. Drilling into the Activity Monitor in the Prompt UI, one can get a detailed verdict on the transaction (double-click). Following the yellow highlights above, the prompt was blocked, and the violation is “Sensitive Data”. The specific offending content, the New York City street address, is flagged as a precluded entity type of “mailing address”. Other fields that might be potentially blocking candidates with Prompt’s solution include various international passports or driver’s license formats, credit card numbers, emails, and IP addresses, to name but a few. A nice, time saving feature offered by the Prompt Security user interface is to simply choose an individual security framework of interest, such as GDPR or PCI, and the solution will automatically invoke related sensitive data types to detect. An important idea to grasp: The solution from Prompt is much more nuanced and advanced than simple REGEX; it invokes the power of AI itself to secure customer journeys into safe AI usage. Machine learning models, often transformer-based, have been fine-tuned and orchestrated to interpret the overall tone and tenor of prompts, gaining a real semantic understanding of what is being conveyed in the prompt to counteract simple obfuscation attempts. For instance, using printed numbers, such as one, two, three to circumvent Regex rules predicated on numerals being present - this will not succeed. This AI infused ability to interpret context and intent allows for preset industry guidelines for safe LLM enforcement. For instance, simply indicating the business sector is financial will allow the Prompt Security solution to pass judgement, and block if desired, financial reports, investment strategy documents and revenue audits, to name just a few. Similar awareness for sectors such as healthcare or insurance is simply a pull-down menu item away with the policy builder. Source Code Detection A common use case for LLM security solutions is identification and, potentially, blocking submissions of enterprise source code to LLM services. In this scenario, this small snippet of Python is delivered to the Prompt service: def trial(): return 2_500 <= sorted(choices(range(10_000), k=5))[2] < 7_500 sum(trial() for i in range(10_000)) / 10_000 A policy is in place for Python and JavaScript detection and was invoked as hoped for. curl --request POST \ --url https://llmsec.busdevf5.net/v1/chat/completions \ --header 'authorization: Bearer sk-oZU66yhyN7qhUjEHfmR5T3BlbkFJ5RFOI***********' \ --header 'content-type: application/json' \ --header 'user-agent: vscode-restclient' \ --data '{"model": "gpt-3.5-turbo","messages": [{"role": "user","content": "def trial():\n return 2_500 <= sorted(choices(range(10_000), k=5))[2] < 7_500\n\nsum(trial() for i in range(10_000)) / 10_000"}]}' Content Moderation for Interactions with LLMs One common manner of preventing LLM responses from veering into undesirable territory is for the service provider to implement a detailed system prompt, a set of guidelines that the LLM should be governed by when responding to user prompts. For instance, the system prompt might instruct the LLM to serve as polite, helpful and succinct assistant for customers purchasing shoes in an online e-commerce portal. A request for help involving the trafficking of narcotics should, intuitively, be denied. Defense in depth has traditionally meant no single point of failure. In the above scenario, screening both the user prompt and ensuring LLM response for a wide range of topics leads to a more ironclad security outcome. The following demonstrates some of the topics Prompt Security can intelligently seek out; in this simple example, the topic of “News & Politics” has been singled out to block as a demonstration. Testing can be performed with this easy Curl command, asking for a prediction on a possible election result in Canadian politics: curl --request POST \ --url https://llmsec.busdevf5.net/v1/chat/completions \ --header 'authorization: Bearer sk-oZU66yhyN7qhUjEHfmR5T3Blbk*************' \ --header 'content-type: application/json' \ --header 'user-agent: vscode-restclient' \ --data '{"model": "gpt-3.5-turbo","messages": [{"role": "user","content": "Who will win the upcoming Canadian federal election expected in 2025"}],"max_tokens": 250,"temperature": 0.7}' The response, available in the Prompt Security console, is also presented to the user. In this case, a Curl user leveraging the VSCode IDE. The response has been largely truncated for brevity, fields that are of interest is an HTTP “X-header” indicating the transaction utilized the F5 site in Toronto, and the number of tokens consumed in the request and response are also included. Advanced LLM Security Features Many of the AI security concerns are given prominence by the OWASP Top Ten for LLMs, an evolving and curated list of potential concerns around LLM usage from subject matter experts. Among these are prompt injection attacks and malicious instructions often perceived as benign by the LLM. Prompt Security uses a layered approach to thwart prompt injection. For instance, during the uptick in interest in ChatGPT, DAN (Do Anything Now) prompt injection was widespread and a very disruptive force, as discussed here. User prompts will be closely analyzed for the presence of the various DAN templates that have evolved over the past 18 months. More significantly, the use of AI itself allows the Prompt solution to recognize zero-day bespoke prompts attempting to conduct mischief. The interpretative powers of fine-tuned, purpose-built security inspection models are likely the only way to stay one step ahead of bad actors. Another chief concern is protection of the system prompt, the guidelines that reel in unwanted behavior of the offered LLM service, what instructed our LLM earlier in its role as a shoe sales assistant. The system prompt, if somehow manipulated, would be a significant breach in AI security, havoc could be created with an LLM directed astray. As such, Prompt Security offers a policy to compare the user provided prompt, the configured system prompt in the API call, and the response generated by the LLM. In the event that a similarity threshold with the system prompt is exceeded in the other fields, the transaction can be immediately blocked. An interesting advanced safeguard is the support for a “canary” word - a specific value that a well behaved LLM should never present in any response, ever. The detection of the canary word by the Prompt solution will raise an immediate alert. One particularly broad and powerful feature in the AI firewall is the ability to find secrets, meaning tokens or passwords, frequently for cloud-hosted services, that are revealed within user prompts. Prompt Security offers the ability to scour LLM traffic for in excess of 200 meaningful values. Just as a small representative sample of the industry’s breadth of secrets, these can all be detected and acted upon: Azure Storage Keys Detector Artifactory Detector Databricks API tokens GitLab credentials NYTimes Access Tokens Atlassian API Tokens Besides simple blocking, a useful redaction option can be chosen. Rather than risk compromise of credentials and obfuscated value will instead be seen at the LLM. F5 Positive Security Models for AI Endpoints The AI traffic delivered and received from Prompt Security’s AI firewall is both discovered and subjected to API layer policies by the F5 load balancer. Consider the token awareness features of the AI firewall, excessive token consumption can trigger an alert and even transaction blocking. This behavior, a boon when LLMs like the OpenAI premium GPT-4 models may have substantial costs, allows organizations to automatically shut down a malicious actor who illegitimately got hold of an OPENAI_API key value and bombarded the LLM with prompts. This is often referred to as a “Denial of Wallet” situation. F5 Distributed Cloud, with its focus upon the API layer, has congruent safeguards. Each unique user of an API service is tracked to monitor transactional consumption. By setting safeguards for API rate limiting, an excessive load placed upon the API endpoint will result in HTTP 429 “Too Many Request” in response to abusive behavior. A key feature of F5 API Security is the fact that it is actionable in both directions, and also an in-line offering, unlike some API solutions which reside out of band and consume proxy logs for reporting and threat detection. With the automatic discovery of API endpoints, as seen in the following screenshot, the F5 administrator can see the full URL path which in this case exercises the familiar OpenAI /v1/chat/completions endpoint. As highlighted by the arrow, the schema of traffic to API endpoints is fully downloadable as an OpenAPI Specification (OAS), formerly known as a Swagger file. This layer of security means fields in API headers and bodies can be validated for syntax, such that a field whose schema expects a floating-point number can see any different encoding, such as a string, blocked in real-time in either direction. A possible and valuable use case: allow an initial unfettered access to a service such as OpenAI, by means of Prompt Security’s AI firewall service, for a matter of perhaps 48 hours. After a baseline of API endpoints has been observed, the API definition can be loaded from any saved Swagger files at the end of this “observation” period. The loaded version can be fully pruned of undesirable or disallowed endpoints, all future traffic must conform or be dropped. This is an example of a “positive security model”, considered a gold standard by many risk-adverse organizations. Simply put, a positive security model allows what has been agreed upon through and rejects everything else. This ability to learn and review your own traffic, and then only present Prompt Security with LLM endpoints that an organization wants exposed is an interesting example of complementing an AI security solution with rich API layer features. Summary The world of AI and LLMs is rapidly seeing investment, in time and money, from virtually all economic sectors; the promise of rapid dividends in the knowledge economy is hard to resist. As with any rapid deployment of new technology, safe consumption is not guaranteed, and it is not built in. Although LLMs often suggest guardrails are baked into offerings, a 30-second search of the Internet will expose firsthand experiences where unexpected outcomes when invoking AI are real. Brand reputation is at stake and false information can be hallucinated or coerced out of LLMs by determined parties. By combining the ability to ingest globally at high-speed dispersed users and apply a first level of security protections, F5 Distributed Cloud can be leveraged as an onboarding for LLM workloads. As depicted in this article, Prompt Security can in turn handle traffic egressing F5’s distributed HTTPS load balancers and provide state-of-the-art AI safeguards, including sensitive data detection, content moderation and other OWASP-aligned mechanisms like jailbreak and prompt injection mitigation. Other deployment models exist, including deploying Prompt Security’s solution on-premises, self-hosted in cloud tenants, and running the solution on Distributed Cloud CE nodes themselves is supported.
