We’re excited to announce that Upwind AI DR (AI Detection & Response) is now generally available.
AI DR gives security teams real-time detection and response for the AI agents running in production, not by bolting on a new sensor, but by extending the runtime intelligence Upwind already has. The Upwind Cloud & AI Security Platform already sees every workload, network flow, identity, and data access across all your cloud providers, and every AI agent and tool across your AI stack, no matter where they’re running.
The AI DR is now able to tell you in real time what an AI agent is actually doing.
Inventory tells you what’s there. Runtime tells you what it’s doing
AI Inventory is the foundation for AI security, you cannot secure your AI Stack until you are fully aware of what is running on your environment, what models are deployed, which agents, tools, mcps and skills are being used.
But an inventory alone is just a snapshot. It tells you what’s there, not what it’s doing right now. Two agents can look identical on paper, same framework, same model, same permissions, and one of them is acting exactly as expected while the other has just been instructed to read data it was never meant to touch.
Is your AI inventory a list, or a live picture?
Agents are built to act, not just run. They call tools, reach APIs, read and write data, and make decisions at runtime, often with credentials and permissions broad enough to get real work done. That autonomy is exactly what makes them useful, and it’s exactly why seeing what an agent does matters as much as knowing it exists. When an agent’s actions shift, whether from a manipulated input, an unexpected tool call, or a session acting differently than usual, that shift is visible the moment it happens, not days later during a review.
That’s the gap AI DR closes, not by replacing inventory, but by stitching it directly into Upwind’s runtime layer, so every asset already sitting in your AI inventory carries a live record alongside it.
Harnessing AI Inventory for AI Runtime Intelligence
Upwind already builds context across the entire platform by correlating network topology, workload identity, API traffic, and data flows into one live picture of the environment. AI DR extends that picture with a dedicated enrichment layer purpose-built for agent behavior:
- Agent Intent tracking: is the user trying to shift the agent into doing something it is not supposed to do? Access tools that are out of his scope? Leverage permissions for supposed restricted actions?
- AI usage detection: identifying where agents and AI frameworks are actually running, inferred directly from runtime traffic, not from a manifest.
- Agent traffic mapping: which tool each agent calls, what data, data stores and APIs it touches, and what a normal session looks like for it.
- Realtime enrichment: when something looks wrong, Upwind attaches the context a responder needs immediately: Identity, data sensitivity, posture and exposure (is this agent over-permissioned or internet-facing?), and the specific action that triggered the detection, instead of a bare alert.
The result is a detection that arrives already answered: which agent, what it did, what it touched, and why it matters, delivered to the customer in real-time rather than surfaced during a retroactive investigation.
Seeing It In Action
How It’s Tracked
Upwind continuously baselines each agent’s behavior, using the same runtime sensors that already monitor workloads across the environment: what it calls, what it accesses, and how often. When an agent’s session activity, tool calls, or destination endpoints drift from that baseline, AI DR flags the deviation as it happens.
How the Agent Was Attacked
Rather than stopping at “something looks unusual,” AI DR reconstructs the path that led there, tracing the entry point back through the agent’s inputs (a tool response, a retrieved document, an upstream prompt) to identify the technique used to manipulate it.
What Data Was Exposed
Because Upwind already maps data flows and sensitivity across the environment, AI DR can immediately show what the agent actually touched once compromised, and which of that was sensitive, so the response team isn’t left guessing at blast radius.
What You Can Do About It
Every AI DR finding ships with response guidance attached: which session or credential to revoke, which permissions to scope down, and which downstream systems to check, so remediation starts the moment the detection lands instead of after a triage call.
Key Takeaways
- Upwind AI DR is now GA, giving security teams realtime detection and response for AI agents running in production.
- It’s built on Upwind’s existing runtime intelligence, network topology, workload identity, posture, and data flows, extended with a dedicated enrichment layer for agent behavior.
- Every detection arrives enriched with the attack path, the data exposed, and recommended remediation, not just a bare alert.
