XNODE Inc.

Xnode today announced Cortx Sentinel, a new capability of the Xnode Cortx AI Control Plane that checks AI agent actions before they run.   Read the announcement: https://shorturl.at/yVXFj Cortx Sentinel is included at no additional cost. Xnode Cortx is in production at asset management and fintech customers and is now available on the AWS, Azure and Google Cloud marketplaces. See Cortx Sentinel at Money20/20 USA, Startup Hub pod SU38. Firms can try it free for 30 days in their own environment, with their own agents.   #Money2020 #AIAgents #AIControlPlane #Fintech

Your bank has an AI budget. Can it say what one agent decision costs?   In a pilot, a few users run one agent. In production, agents can call models, tools, and other agents. The bill can grow fast.   Money20/20 has many sessions about moving beyond pilots. One focuses mainly on cost: “Making AI Pay: What It Really Costs to Scale Beyond the Pilot,” Monday at 3:35 PM.   A business decision can involve several agents and models. Without cost attribution, their spending can disappear into one total.   AI FinOps looks at the cost of finishing each workload, which model handled it, how old that deployment is, and how much context or state it is carrying.   You cannot judge the return on spend you cannot trace to an agent and its owner.   Xnode Cortx attributes every token to a user, a team and an agent. Alerts fire before a budget limit, and spend stops at it.   If your CFO asked what your busiest agent cost last week, how long would the answer take?   Meet Kumar Velugula at Startup Hub Pod SU38, The Venetian, October 18 to 21.   #Xnode #AIControlPlane #AIFinOps #Money2020

One Money20/20 panel calls it unsettled: who decides when an agent may start something on its own?   About a dozen sessions at Money20/20 touch on agent accountability. Most focus on outside agents that shop and make payments. Fewer ask about the agents a bank runs itself.   The panel “Is Acting a Responsibility Too Far for Agents?” leaves two questions open. When can an agent act on its own? Who takes responsibility when it does?   Identity tells you who the agent is. The workflow decides when it can act alone.   Each step needs a clear boundary: can the agent proceed, or must a named person take over?   Agent orchestration decides which agent handles each step, in what order, and when the work stops.   Which agent decision would you never let run without a named person ready to step in?   Meet Kumar Velugula at Money20/20 USA, Startup Hub, Pod SU38, The Venetian, October 18 to 21.   #Xnode #AIControlPlane #AgentOrchestration #Money2020

On April 17, the Fed, OCC and FDIC replaced SR 11-7 with SR 26-2. One footnote matters more than the rest: generative and agentic AI are out of scope. That doesn’t leave agents ungoverned. Fair lending, third party risk, cybersecurity and recordkeeping rules still apply, and the agencies plan to ask the industry for input on AI next. Model risk management assumed a model changes through a reviewed release. An agent changes without one. Someone edits a prompt on Tuesday. A new tool gets added Wednesday. The vendor updates the model overnight. What you validated in March may not be what’s running today. So the evidence has to come from runtime: which rule applied to which action, under which version of the policy. The policy also has to live outside the agents. More guidance is coming, and if every rule change means rebuilding every agent, each change costs more with every agent you add. That’s why we built Xnode Cortx, our AI Control Plane platform, so policy is enforced on every request and changes without redeploying a single agent. Pico runs it in production today. Next week at Money20/20 USA, regulation and AI supervision sit under the “Brave New World: Setting Guardrails” pillar. Our question for the room: What would you show an examiner today as proof that your agents followed your own rules? If you’re working through the same question, find Kumar Velugula at Startup Hub Pod SU38, The Venetian, October 18 to 21. #Money2020 #AgenticAI #ModelRisk #AIControlPlane

What if the AI tokens price reduces, what do you think happens to your AI bill? Does it go down, or go up? Accenture researched 750 senior executives from 17 countries and released the report on AI Tokenomics. As per the report, only 1 in 20 executives would bank the savings if the token prices were reduced by 25% or more. The rest of the organizations will likely expand the AI usage. But why? The reason being that when the AI becomes cheaper, the companies start using AI, even for tasks that once cost too much. The cost per task may be reduced, but the total usage might increase. This is why lower token prices do not automatically mean lower AI bills. Accenture calls it a Jevons paradox. When steam engines became more efficient, Britain ended up consuming more coal, not less. AI could follow the same pattern. There is one more problem. When an enterprise looks at the bill from an AI vendor, they cannot figure out what tasks the AI was used most, or which user used how much, or for how much spend they got how much output.  One case mentioned in the report was one of the Asian banks. They started using AI coding assistants and after three months the AI bill was way over then the forecast. The engineers were using the most expensive model for every type of task. Instead of using expensive models for complex tasks and cheaper models for day-to-day routine tasks. They had no way to know it. There is a way by which a company can learn about different workload costs, model usage, deployment time, and how much context it is carrying. It is called AI FinOps. It is important to understand that that token price is an input cost, but businesses need to learn the actual cost for completing the workload. Do you know the AI cost of each finished task in your enterprise? #Xnode #AIControlPlane #AIFinOps #EnterpriseAI #AIGovernance

Every company has an AI data access policy, but when an AI agent accesses that data, is the policy checked at that exact moment? Delinea has published a 2026 Identity Security report. Organizations with more than 500 employees from eight different countries were surveyed. Almost all organizations had a formal policy that defined what data AI tools and agents can access. But only half of them check access in real time whether the data access by AI agents is as per the policy. For example, a company clearly writes in its policy document what type of data can be accessed by the AI agents and what the permissions are. Now the issue is that the policy is not enforced at the time when the agent is accessing the data. Policy is checked at the time of review, audit, or login. A policy existing only in documents has no meaning, as it cannot control the actual action. This is where policy and governance under an AI Control Plane comes in handy, which means policy stated once, applied at every request and changed without redeploying the systems it governs.   As per the Delinea report, policy enforcement is not consistent across different environments. Policy applies in some environments and not in others. The weakest environments are CI/CD pipelines and Kubernetes. Picture this: a bank has a policy that no customer data will reach unauthorized external AI models. This rule is checked at the data warehouse, but a coding agent runs in the CI/CD pipeline. The agent got blocked on the warehouse data route, but it got the data through the pipeline. When a policy applies inconsistently across environments, the weakest environment policy becomes the actual policy. Wherever the enforcement is missing, the agent finds the gap. When we say, "applied at every request", we literally mean every environment. It includes every application, warehouse, pipeline, Kubernetes, etc. Wherever the agent requests data, the policy applies. Whether the request comes through the customer data warehouse or CI/CD pipeline, the rule should stay the same.  In which environment is your AI policy weaker? #Xnode #AIControlPlane #PolicyAndGovernance #EnterpriseAI #AIGovernance

If your AI agents can talk to each other, who decides whether one can hand work to another?   Mass General Brigham's Chief Information and Digital Officer, Jane Moran, told Becker's Healthcare that agentic AI is pushing her toward a more restrictive stance sooner than she expected. MGB has spent three years building AI governance, so this isn't a newcomer's caution.   She described a scenario. One agent has a narrow job, say password resets. Another agent elsewhere in the system recognizes that role and could, in theory, ask it for access across systems. So for now, MGB isn't allowing autonomous AI across platforms.   The detail that stands out: part of the reason traces back to the CrowdStrike outage, when MGB found systems on its network that weren't in its configuration management database. You can't govern what you can't see. Agents make that harder, because they create new connections at runtime.   Suppose the scenario plays out. Check each step on its own and nothing fails. One agent asks; the other answers; every action sits inside its permissions. The risk only exists in the chain, and most governance tools still judge one step at a time.   At XNODE we think two principles get enterprises from blanket restriction to safe delegation.   Authority travels with the task. When one agent asks another for something, the request should carry the authority of the person who started the chain, and no more. Delegation can narrow access; it should never widen it.   Autonomy is earned. Moran expects agents to get closer to employee-level autonomy, with each expansion tied to the risk profile of the task. We'd add the agent's own record. An agent that shows its plan before it acts, and keeps to it, earns more room. An agent that hides its plan, or once tried to work around a limit, gets less.   Both need one place where every agent is known and every handoff is visible before it runs. That's the job of an AI Control Plane.   When one of your agents asks another for help, whose authority is it using?   #Xnode #AIControlPlane #AgentOrchestration #AgentGovernance #HealthcareAI #AIGovernance

Do you know all the AI agents running in your org? Do you know what AI tools are being accessed by employees? According to Gartner, 75% of organizations reported unauthorized use of AI coding assistants. Gartner has put agentic AI oversight in its Top Trends in Cybersecurity for 2026. Agentic AI is expanding rapidly, and many organizations are not able to keep up with the growth of agents and the challenge of identifying and governing them. There are multiple ways a shadow agent can come inside the enterprise. It can arrive embedded in enterprise software, come directly from the internet, or be built by employees.  This is the shadow AI: agents and tools running in a business without IT's knowledge.  What is an easy solution here? The easy answer is to list every agent. Write the permissions and scope document for each of them. But this process might not be fast enough. If the agent is in shadow, then how would an organization know about its existence? Without knowing the agent, how can we make the list? Gartner suggests a "track back" inventory. Instead of just finding the agents, start monitoring changes in utilization patterns for sensitive systems and data activity. Imagine a bank that tracks usage on its customer-record API and the payments API. If the bank sees an unusual change in the API usage, the bank can trace it back and find out the following things: - Which application requested the call?  - Which agent initiated the request?  - Is the agent registered or not? If agents can access the tools, APIs, and enterprise data, then just keeping an eye on registered agents is not enough. The layer that sits on the data request point can see all agents' requests that pass through it, regardless of whether that agent is listed or not. That conceptual control point is called the AI Control Plane. As the number of agents increases inside the enterprise, it is important that those agents are discovered at the same pace using data and systems activity. Ultimately, detecting shadow AI is needed to mitigate enterprise risk. #Xnode #AIControlPlane #ShadowAI #EnterpriseAI #AIGovernance

Last year at Money20/20 USA, a Citi panel with Mastercard and Consult Hyperion asked a simple question: when AI acts independently, who bears responsibility? A year on, AI agents are preparing and executing real steps across financial workflows. The question is still open. We think part of the answer starts with a question few teams are asking yet. Should every agent get the same permissions? A payments agent that shows its plan before it acts, and has never tried to work around a limit, has earned something an opaque agent hasn't. What that looks like in practice is a conversation we'd rather have in person.   Xnode founder Kumar Velugula will be present at Money 20/20 USA 2026 in Las Vegas to talk about similar questions. When an AI agent takes action in the financial workflow, then at that moment who decides its identity and permissions?   Event Details:  Money 20/20 USA  The Venetian, Las Vegas  October 18-21, 2026  Startup Hub, Pod SU38   If you are attending the event, please stop by our Pod SU38.   You can also send us a request through the Money 20/20 Connect App.   #Money2020 #Xnode #AIControlPlane #AgenticAI #FinancialServices

What if you are one of those whose images were leaked by ChatGPT to a website? OpenAI recently disclosed another security incident. Given its track record, it felt more like a scheduled update than some news. Its agent leaked 53 images of ChatGPT users to some external image hosting website. The images were shared as unlisted links. The incident is not a hack, since the images belonged to consumers who had not opted out of having their content used for model training. But the agent went rogue and took an unpredicted action.  In another incident, the OpenAI models interacted with US government websites in an unexpected way. In an earlier report, an OpenAI agent hacked an Australian government health website, gaining unauthorized access to some files. Things seem to be getting out of control! Literally. All these incidents tell us one thing: that an agent's behavior cannot be completely controlled. An agent is given access to different tools and data to perform certain actions. To execute the tasks, sometimes an agent takes an unintended path which is hard to predict. It is fine to have agent autonomy, so that it has the freedom to do tasks in its own way, but there should be a set of boundaries which it cannot cross.  It is important that every agent should have its own identity and access as the employees do. Only essential permissions should be granted to an agent. So even if it tries to do an unexpected action, the limited permissions can reduce potential damage. It is not enough to have access permissions defined well. We need policies and guardrails that permit actions, block restricted actions, and require approval for sensitive actions. After we have the preventive controls in place, we would also need the full-stack observability to keep an eye on what agents are doing while they are running. Every small step taken by the agent needs to be tracked.  We also need an emergency control system to intervene when an agent goes rogue. Something like an Agent Control Tower, from which we can suspend or stop an agent immediately. We need multiple capabilities to run a smooth operation. We cannot get away with just having identity and access or just policy and governance. When we bring all these capabilities under one umbrella, we call it an AI Control Plane. It is non-negotiable for enterprises to have a control plane between their models and their data. Enterprises need to control what an AI agent can access, what it can do, and how quickly they can intervene when it acts beyond its intended task. Who should be responsible for stopping an AI agent when it acts beyond its intended task? #AIControlPlane #EnterpriseAI #AIGovernance #AIAgents #AIObservability