The gap
A backlog is not a security posture
Ask a cloud security team how many open findings they have and you will hear a number in the thousands. Ask how many an attacker could reach and exploit today and you will get a pause, because the tooling was never built to answer that question.
We spent a decade building systems that are very good at describing what could go wrong in a cloud environment, and almost none that show what is actually happening inside it. Most teams have been solving the first problem and reporting it as progress on the second.
Posture tells you what could happen. Runtime tells you what is happening, and only runtime can prioritize the other.
The problem
What a configuration snapshot cannot see
Agentless posture tools were a real advance. They read your cloud configuration through APIs, compare it against a policy library and produce findings, with no software on the workload and coverage in an afternoon.
The trouble is what they can see. A critical CVE in a library that no process ever loads gets the same score as one that is loaded, internet-facing and running with access to your production database, because from the outside the two look identical. Multiply that across an environment and you get a backlog nobody can finish, which teams eventually stop reading.
Swipe to see the full diagram
Upwind’s founders came to this problem from DevOps rather than security. Their previous company, Spot.io, was built on understanding what cloud workloads actually do, and NetApp acquired it in 2020.
The platform
One sensor, one source of truth
Upwind deploys an eBPF sensor into each workload. eBPF lets the Linux kernel safely observe process execution, system calls, network connections and file access, without modifying the kernel and without the overhead that made older agents unwelcome. You get a continuous view of behavior in place of a scheduled scan.
That view changes the ranking. A vulnerable package that never loads into memory drops down the list, and one that is loaded, internet-facing and running with write access to a production bucket rises to the top with its attack path attached. The CVSS scores are identical, and observation is what separates them.
Upwind runs posture, application security and real-time protection off that same sensor, covering categories that used to take half a dozen vendors. Each one improves when they share live context, because prioritization is only as good as the evidence underneath it.
Swipe to see the full diagram
The objection
I already run a CNAPP. Why does this change?
Most enterprises finished rolling out a cloud-native application protection platform in the last couple of years and have no appetite to replace it, so this comes up in almost every conversation we have about runtime.
Coverage and context are different capabilities, and your CNAPP almost certainly gave you the first one.
Most incumbent platforms were built agentless-first and added runtime later as a module, with a second sensor and data model feeding a scoring engine designed to work without them. That is why the backlog survives deployment. Ephemeral infrastructure widens the gap, since containers, serverless functions and AI agents exist only while they run and leave little configuration to snapshot.
For most teams the decision is which layer becomes the source of truth for prioritization, and the time to make it is before the AI stack makes it for you.
What’s next
Your AI agents are workloads too
An AI agent makes its decisions at runtime. It calls tools, invokes models, reads data and hands work to other agents, and a configuration snapshot of that system tells you almost nothing about which calls it made or which data it touched.
In June, Upwind was one of the first security platforms to integrate with Cisco Cloud Control at launch, bringing its runtime findings, API telemetry and AI visibility into Cisco AI Canvas over the Model Context Protocol. The vendors building the agentic operations layer needed a source of runtime truth for AI workloads, and they went to a company whose whole architecture is runtime observation.
In September, Upwind acquired Aegis, a startup focused on securing AI agents, and launched Upwind AI Labs for Cybersecurity to extend the same runtime view to the agents themselves.
Proof
What Upwind customers report

Vectra AI’s security team had full visibility into its containers within hours of deploying Upwind.
Read the customer story
CAVA went from about a dozen false positives a day to none, so every alert that fires now gets taken seriously.
Read the customer story
Sprinklr’s product security team chose Upwind for its view across the full stack, application layer included.
Read the customer story
H2O.ai had Upwind running in production for all of its customers in five minutes, closing the gap between security and DevOps.
Read the customer story