Agentless scanning and eBPF runtime sensors in one CNAPP
Upwind is a cloud-native application protection platform (CNAPP) for security, DevOps and platform teams running workloads in public cloud.
AI Panel Score
6 AI reviews
Reviewed
AI Editor ApprovedApproved and published by our AI Editor-in-Chief after full panel analysis.Upwind is a cloud-native application protection platform that pairs agentless cloud scanners with eBPF runtime sensors, so posture findings carry evidence of what is actually running, reachable and exposed. It targets security, DevOps and platform engineering teams operating production workloads on AWS, Azure, Google Cloud or Oracle Cloud Infrastructure. Pricing is quote-based: upwind.io publishes no price list and routes buyers to a demo, while the vendor's AWS Marketplace listing offers a free trial plus 12-, 24- and 36-month contracts. Core capabilities include function-level reachability that confirms whether a vulnerable package is actually called, attack-path analysis across identities and workloads, discovery and schema mapping of REST, GraphQL, gRPC and SOAP APIs with PII, PHI and PCI classification, and an Agentic Pack whose Blue, Red and Green agents investigate, validate and remediate findings under a conversational control plane called ChoppyAI. It best fits teams consolidating posture, workload protection and detection and response into one platform.
Getting started means connecting a cloud account - AWS, Azure, Google Cloud or Oracle Cloud Infrastructure - and installing the eBPF sensor on the workloads you want covered. Cloud Scanners either run in Upwind's own tenancy (the SaaS model) or inside your accounts (the Outpost model), and sensors ship for Kubernetes, Linux and Windows Server hosts, Amazon ECS on EC2 and Fargate, AWS Lambda, Azure Container Apps and Google Cloud Run. From there the Management Console shows a live inventory and asset graph, with work organized into Code, Configurations, Attack Surface, Identities, Vulnerabilities, Threats, API Security, Data Security and AI sections.
The runtime layer is what separates findings that matter from the rest. Function-level reachability checks whether a vulnerable package is actually called before a CVE is escalated, attack-path analysis chains individually minor findings into the routes an attacker could take, and the platform baselines normal cloud and workload behavior so privilege escalation, lateral movement and unusual API use surface as deviations. API security discovers managed, unmanaged and shadow APIs across REST, GraphQL, gRPC and SOAP, maps schemas from live traffic, and classifies PII, PHI, PCI and secrets moving through them. The Agentic Pack adds four named AI agents: Blue investigates incidents from first signal to verdict, Red probes the live environment to prove which risks are reachable, Green groups findings into single high-impact fixes, and ChoppyAI answers plain-language questions across the whole data model.
It is aimed at security, DevOps and platform engineering teams already running production workloads in public cloud who want posture management, workload protection, identity entitlements, data security and detection and response in one platform rather than five. Pricing is quote-based: upwind.io lists no prices and routes buyers to a demo request, while the AWS Marketplace listing offers a free trial alongside 12-, 24- and 36-month contracts and private offers. A 24/7 Managed Detection and Response service is sold as a separate line.
The integration surface is wide. There is a documented REST API (v1 and v2), a Terraform provider that manages detection policies and access scopes declaratively, an upwindctl CLI for scanning files, images and machine images, a hosted MCP server at mcp.upwind.io for MCP-capable AI clients, and SAML single sign-on with Okta, Microsoft Entra ID, Google Workspace, Duo, JumpCloud, OneLogin and PingOne. Outbound integrations cover Jira, Slack, Microsoft Teams, ServiceNow, Linear, PagerDuty, Splunk, Datadog, Tines and custom webhooks, and CI/CD connectors cover GitHub Actions, GitLab, CircleCI, Argo CD, Bitbucket and Harbor.
Blue investigates incidents from first signal to verdict, Red probes the live environment to prove which risks are reachable, Green consolidates findings into single high-impact fixes, and ChoppyAI answers plain-language questions and dispatches the other three.
Auto-discovers managed, unmanaged and shadow APIs across REST, GraphQL, gRPC and SOAP and builds their schemas from observed runtime traffic rather than from source annotations.
A Terraform provider manages detection policies, custom rules, access scopes and malware indicators declaratively through the Management REST API v2, alongside a documented v1 and v2 REST surface and ten published onboarding and cloud-scanner modules for AWS, GCP, Azure and OCI. The provider is early — v0.1.x, community tier — so the modules carry the mature onboarding path.
Finds and classifies PII, PHI, PCI data and secrets in cloud data stores and in traffic moving through APIs, and flags when regulated data crosses an unsafe boundary.
Cloud Scanners inventory AWS, Azure, Google Cloud and Oracle Cloud Infrastructure accounts without an agent, running either in Upwind's tenancy (SaaS model) or inside your own accounts (Outpost model).
Maps permissions from identity providers through to cloud resources, tiers privilege levels, and surfaces dormant roles, unused access and exposed secrets.
A hosted Model Context Protocol server at mcp.upwind.io lets Claude, Cursor and other MCP clients query Upwind security data over OAuth 2.1, with regional endpoints such as mcp.eu.upwind.io.
A Kubernetes admission controller blocks deployments carrying exploitable vulnerabilities, malware or exposed secrets, with policy scoped by namespace, cluster and workload type.
Continuously evaluates cloud configuration against built-in and custom rules written in Rego, maps findings to compliance frameworks and produces audit-ready reporting.
Chains individually minor findings into the routes an attacker could take across workloads, identities and data, and shows the blast radius of each path.
A lightweight eBPF sensor installs on Linux and Windows Server hosts, Kubernetes nodes, Amazon ECS on EC2 and Fargate, AWS Lambda, Azure Container Apps and Google Cloud Run to stream live process, network and syscall telemetry.
Builds SBOMs at both build time and runtime across container images, machine images, IaC templates and repositories, with an SBOM Explorer for tracing a package to the workloads that run it.
Baselines normal cloud and workload activity, then flags deviations such as privilege escalation, lateral movement and unusual API calls and correlates them with audit logs into a single incident timeline.
Runtime call-graph mapping checks whether a vulnerable function is actually invoked in production, so unreachable CVEs are ranked below ones an attacker can use.
Core CNAPP platform. Listed at $30,000 per unit for a 12-month contract on Upwind Security's own AWS Marketplace listing. Longer commitments are advertised only as ceilings — "save up to 8%" on 24 months and "save up to 11%" on 36 months — with no dollar figure published. upwind.io itself publishes no price list.
24/7 managed detection and response, billed as a separate line. Listed at $6,000 per unit for a 12-month contract on Upwind Security's own AWS Marketplace listing. Longer commitments are advertised only as ceilings — "save up to 8%" on 24 months and "save up to 11%" on 36 months — with no dollar figure published.
Custom-quoted terms. upwind.io routes all buyers to a demo request and publishes no prices; AWS Marketplace offers a private-offer path for negotiated volume and contract length. A free trial is available through the marketplace listing.
The runtime layer earns the $30,000, but you'll only find that number on AWS Marketplace.
“Upwind pairs agentless cloud scanners with an eBPF sensor, so findings carry proof of what is actually running. The platform is deep, but the price you can look up sits on a marketplace listing, not on upwind.io.”
Upwind won't quote you on its own site. Every path ends at Get a Demo, and the only rate card I could find is its own AWS Marketplace listing — $30,000 for twelve months of the platform, $6,000 more for the 24/7 MDR line.
What you're buying is runtime evidence. Function-Level Vulnerability Reachability checks whether a vulnerable function is actually called in production before a CVE gets escalated; the obvious posture-only alternative can only tell you the package is installed. That changes which tickets your engineers work on, not just what you pay.
The catch is the term ladder. The discount for committing shows up at 24 and 36 months — up to 8% and 11% off — so you'd sign years before knowing the eBPF sensor covers your ECS, Lambda and Windows Server workloads. Take the free trial, instrument one production cluster, and sign twelve months before anything longer.
Function-level reachability and API schemas built from observed runtime traffic rather than source annotations are things a posture-only scanner cannot produce.
SAML single sign-on across seven identity providers, RBAC, audit logs and an Outpost model that keeps scanners inside your own accounts all cut lock-in, though I couldn't find published cancellation terms.
Connecting an AWS, Azure, Google Cloud or Oracle account gives inventory quickly, but full value waits on rolling the eBPF sensor across Kubernetes, ECS, Lambda and Windows Server.
Posture, workload runtime, identity entitlements, data security and detection land in one console, so it replaces a stack rather than shaving cost off one tool.
Release notes for the sensor, operator and Helm charts, a documented REST API v1 and v2, a Terraform provider and the upwindctl CLI all read as a platform under active build.
Security teams who run production workloads across more than one public cloud.
Small teams who cannot commit five figures a year to cloud security.
Runtime telemetry is where the lock-in lives; scanner placement and Rego policy stay yours.
“Upwind pairs agentless Cloud Scanners with an eBPF runtime sensor, and it will run those scanners inside your own cloud accounts if you want them there. The posture layer is portable policy-as-code; the runtime intelligence that justifies the platform is not.”
Ask what you'd take with you on the way out. Custom posture rules go in as Rego through Upwind Explorer, so that logic stays legible outside their console. The runtime graph doesn't travel: function-level reachability and attack-path chaining come from the eBPF sensor's telemetry model, and that stays with the vendor.
The other decision worth architecture time is where the Cloud Scanners execute. Outpost runs them inside your own AWS Organization, under your SCPs and logging; the SaaS model runs them in Upwind's tenancy. That's a residency decision made at onboarding, and unwinding it later is a re-onboarding project.
Their own People.ai case study sells the switch off Wiz as a coverage story. But upwind.io publishes no prices, and the one rate card they publish themselves — an AWS Marketplace listing at $30,000 per unit for twelve months — never defines the unit. A three-year commit gets negotiated, not modeled.
A runtime-first consolidation play against agentless-first platforms, and their own customer case study names the incumbent they are displacing.
Onboarding runs through Terraform and CloudFormation, custom posture rules accept raw Rego, and detections route into Jira and PagerDuty.
SAML across seven identity providers, six CI/CD connectors, a hosted MCP server at mcp.upwind.io, a documented REST API and the upwindctl CLI.
The Outpost model keeps scanner workloads inside your own AWS Organization, but I couldn't find an export path for the runtime telemetry the platform is built on.
Function-level reachability, attack-path chaining and SBOMs built at both build time and runtime are hard things to do well, and all three are documented.
Security and platform teams who run production workloads across multiple public clouds.
Teams who need a published price list to plan a budget.
Upwind publishes $30,000 for twelve months, per unit — and I couldn't find what a unit counts.
“Upwind's own AWS Marketplace listing prices a 12-month platform contract at $30,000 and the 24/7 MDR line at $6,000. Both are per-unit figures, and I couldn't find what a unit counts, so the published number anchors a budget rather than building one.”
Upwind sells by the unit. I couldn't find what a unit counts. Its own AWS Marketplace listing puts $30,000 on a 12-month platform contract and $6,000 on the 24/7 MDR line. Two numbers, no denominator — two more than the category usually publishes.
At quantity one: platform plus MDR is $36,000 for twelve months. Three annual renewals at list: $108,000 by my arithmetic, before any private-offer discount. The 24- and 36-month terms advertise a discount ceiling but no dollar figure, so twelve months is the only term you can budget.
The $6,000 MDR line is 20% of the platform price — cheap against staffing a 24/7 rotation, and optional. The Agentic Pack pushes the same way: Blue works an incident to a verdict, Green collapses findings into single fixes. Procurement is light. It bills on AWS paper, against commitment you've signed.
Marketplace purchase means existing AWS paper, a free trial before commitment and a documented private-offer route.
Three contract lengths and a private-offer path give real negotiating structure, though only the 12-month figures are published.
Upwind's own AWS Marketplace listing publishes $30,000 for a 12-month contract, but upwind.io publishes nothing and I couldn't find what a unit counts.
The Agentic Pack and Attack Path Analysis produce countable outputs, and one $30,000 line replaces CSPM, CIEM, DSPM and vulnerability-management renewals you can diff against.
Platform plus MDR runs $36,000 a year at quantity one, though Outpost-model scanners put scanning compute on your own cloud bill.
AWS Marketplace buyers who want cloud security on existing cloud paper.
Budget owners who need a modelable unit price before a sales call.
Runtime proof a vulnerable function actually ran — once you turn the JVM and Go hooks on
“Function In-Use uses eBPF uprobes to confirm a vulnerable function was actually executed, which is the difference between a CVE list and a work queue. The per-language hooks on hosts ship mostly off, so the first week is a config pass, not a dashboard tour.”
The CLI flag table has a column headed "If you omit it." Small thing, tells you who wrote the page. shiftleft scans an image, a directory or an AMI, takes --max-severity to fail a build step, and writes GitLab vulnerability JSON or GitHub job-summary markdown.
Function In-Use does the real work — eBPF uprobes on interpreter internals confirm a vulnerable function actually executed, and Findings has a filter for exactly that. Agentless-only scanners can say a package is present; this says it ran. But on hosts the per-language hooks in agent.yaml default off. Python and native binaries are on; Java, Go, Node.js and PHP are not.
Overhead is governed, not promised: attachment pauses above 15% CPU, capped at 750 processes and 100 hunted functions. Sensor release notes land every couple of weeks. JVM and Go shops should budget a config pass before that signal shows up in triage.
The sensor is eBPF on Linux and ETW on Windows Server 2016 or later, and a Health Center lists unhealthy sensors with 24-hour CPU and memory per host.
The docs run to 643 pages including 22 troubleshooting pages, a flag table with an If you omit it column, and Cluster Manager sizing in CPU cores.
Per-language in-use hooks default off in agent.yaml, and leaving --upwind-uri at its default fails authentication outside the US.
REST API v2 exposes bulk policy create, edit and delete, SBOM Explorer traces a package to the resources running it, and a hosted MCP server sits at mcp.upwind.io.
shiftleft returns GitLab vulnerability JSON and GitHub job-summary markdown, and a Terraform provider manages threat policies and access scopes declaratively.
Platform engineers who triage cloud CVE backlogs every week.
Teams who cannot install a runtime sensor on production hosts.
Sensor installs in one curl line; the console you'll still be learning in month three.
“Connecting a cloud account is a CloudFormation stack and one curl line, and the install docs read like somebody watched people do it. The console behind that door is wide, nine sections deep, and I couldn't find a mobile app for any of it.”
They tell you the client secret shows once and then never again. Small thing, but somebody wrote that line after watching people fumble it. Drop a CloudFormation stack in, run one curl line on your Linux hosts, validate, and you're collecting — as long as the kernel is 4.14 or newer.
The docs keep that up. Release notes are split per component — sensor, operator, Helm charts, upwindctl and six more — and the deprecated three-stack AWS onboarding is still written down instead of quietly deleted. That's a team that expects you to be here next year.
Month three is the wide part. Nine console sections, Code through AI, and ChoppyAI is the escape hatch — plain-language questions instead of hunting menus like the incumbent consoles make you. But I couldn't find a mobile app, and for a product about catching attacks live, the alert wakes you and the laptop answers it.
The install guide flags that the client secret appears only once and keeps the deprecated AWS onboarding path documented rather than deleted.
Nine console sections and a documentation site in the hundreds of pages take time, though ChoppyAI shortens the hunt.
I couldn't find a mobile app, though this is cloud infrastructure tooling where a phone was never really the point.
Connect a cloud account, generate credentials, run one curl line, validate — four steps, though the way in still starts with a demo request.
Spinners, error states and autosave are not things I can judge without sitting in the console, so this one stays neutral.
Platform engineers who need runtime visibility on Kubernetes and Linux hosts.
Teams whose Linux fleet still runs kernels older than 4.14.
93% noise reduction and detection in 15 seconds, with no method published behind either number.
“What Upwind actually ships is unusually concrete — Rego posture rules, a Kubernetes admission controller, API discovery from live traffic. The proof points on the homepage are not.”
The Agentic Pack ships AI agents with names and jobs. Blue investigates, Green consolidates, ChoppyAI answers plain-language questions. Red is the one to think about — it probes your live environment to prove which risks are reachable. Autonomous probing in production is a real capability and a real blast radius.
The homepage leads with 93% noise reduction, detection in 15 seconds, live inventory every 30 seconds. I couldn't find a method behind any of them. Category norm.
Exit is better than I expected. Posture rules are written in Rego, an open policy language, and a Terraform provider keeps detection policy in your repository rather than their console. But the catch is what you cannot take: detection baselines and API schemas are learned from your live traffic, and none of that learned state follows you to another vendor.
Pairing agentless Cloud Scanners with eBPF runtime sensors is a real architectural choice, though every platform in this category now markets runtime context.
Rego policy and a Terraform provider travel with you, but eBPF sensors across every workload and a learned detection baseline do not.
Product posts dated through mid-September 2026, a documented REST API at v1 and v2, a Terraform provider, upwindctl and a hosted MCP server all point to active work.
The site publishes no price and routes every buyer to a demo, while 93% noise reduction and 15-second detection carry no published method.
Named capabilities are concrete — Rego posture rules, function-level reachability, API discovery across REST, GraphQL, gRPC and SOAP — and the documentation backs them.
Platform teams who want runtime proof before escalating a CVE.
Teams who cannot let an AI agent probe production.
Common questions answered by our AI research team
Upwind's AWS Marketplace listing prices the core platform at $30,000 per unit for a 12-month contract and its 24/7 MDR service at $6,000. The upwind.io site publishes no price list and sells through a demo request, with private offers for negotiated terms.
Both. Agentless Cloud Scanners inventory AWS, Azure, Google Cloud and Oracle Cloud accounts, while a lightweight eBPF sensor runs on hosts, Kubernetes nodes, ECS tasks, Lambda and container apps for runtime telemetry.
Yes. Upwind creates and updates Jira tickets from its findings and pushes detections into Slack channels. Other integrations include Microsoft Teams, ServiceNow, Linear, PagerDuty, Splunk, Datadog, Tines and custom webhooks.
Upwind connects AWS, Azure, Google Cloud and Oracle Cloud Infrastructure accounts. Cloud Scanners run either in Upwind's own tenancy under the SaaS model or inside your accounts under the Outpost model, onboarded with Terraform or CloudFormation.
Yes. Upwind documents SAML single sign-on for Okta, Microsoft Entra ID, Google Workspace, Duo, JumpCloud, OneLogin and PingOne, plus a generic SAML option. Role-based access control and audit logs are built into the Management Console.
Company
Upwind SecurityFounded
2022Pricing
Contact for pricingFree Trial
Available
Upwind Security is a San Francisco-based cloud security vendor offering a runtime-powered CNAPP that combines agentless scanning with eBPF sensors.
