The security problem posed by AI agents is not really about the models themselves. It is about everything wrapped around them: the skills, the tools, the connections, and the instructions that determine what an agent actually does once it is set loose in a live environment. Upwind Security is tackling that problem directly with two new releases: the AI Agent Context Scanner and the general availability of AI Detection & Response, or AI DR.

Why Context Matters More Than the Model

AI agents today are functionally different from the applications security teams are used to defending. They call tools. They invoke APIs. They access data stores. They make decisions and execute them using real enterprise credentials, often without a human reviewing each step. And critically, how an agent behaves is shaped by more than its underlying model. It is shaped by a dynamic context assembled in real time from skills, tools, Model Context Protocol connections, memory, and active instructions.

Every one of those components is a potential entry point for something dangerous. A skill can carry a specialized prompt with hidden workflow logic. A tool can be misconfigured to allow more access than intended. An MCP connection can pull in external instructions that were never vetted by anyone on the security team. Because each of these can inject content directly into an agent's context window, a single bad instruction has the power to redirect the agent's entire reasoning process and put its permissions to malicious use.

The Context Scanner's Approach

Upwind's new AI Agent Context Scanner is built specifically to get ahead of that risk. It continuously analyzes the building blocks assembling an agent's operational context: the skills containing prompts and workflow logic, the tools enabling actions like querying databases or running scripts, and the MCP connections funneling in external data, memory, and instructions.

The scanner does not limit itself to a single environment. It evaluates these components wherever they actually run, across employee endpoints, cloud infrastructure, and managed AI providers, looking for unsafe or unintended instructions before an attacker or a careless configuration can exploit them. Once findings surface, they are enriched with additional context from the broader Upwind platform. That step matters because it lets security teams prioritize based on real runtime risk instead of chasing down every alert a scanner produces in isolation.

What Happens After Deployment: AI DR

Scanning context before deployment solves half the problem. The other half is knowing what an agent does once it is running. That is where AI DR comes in, now generally available and extending Upwind's existing runtime intelligence into the world of AI agents specifically.

AI DR continuously builds a behavioral baseline for each agent: which tools it typically calls, which APIs and data stores it normally accesses, and what a standard session looks like. When an agent's tool calls, destination endpoints, or overall session behavior start to drift from that baseline, AI DR flags the deviation as it happens, not after the fact.

From there, Upwind correlates the flagged activity with a wide set of signals, including workload identity, network topology, API traffic, data sensitivity, permissions, posture, and exposure. The output is not a generic alert. It is built to answer the questions a security team actually needs answered: which agent was affected, what did it do, what did it touch, and why does it matter. AI DR can also reconstruct the path that led to the behavior, identify data that may have been exposed, and recommend a response, whether that means revoking a session or credential, reducing permissions, or reviewing downstream systems that may have been touched.

The Bigger Shift: From Inventory to Continuous Security

None of this replaces the need for basic AI inventory. Organizations still cannot secure agents, tools, skills, or MCP connections they do not know exist, and that discovery layer remains foundational. But Upwind's broader argument is that inventory by itself is only a snapshot. Two agents can appear identical sitting in a catalog and behave in entirely different ways once they are actually running. By linking preventive context scanning with real-time detection and response, Upwind is aiming to give security and engineering teams a continuous, live view of the instructions shaping AI agents and the actions those agents take.

What the Release Includes

**Taken together, the two capabilities span the full lifecycle. On the discovery side, Upwind identifies and analyzes agents, tools, skills, and MCP connections wherever they live, whether on endpoints, in cloud infrastructure, or with managed AI providers. On the analysis side, it digs into the instructions injected into each agent's operational context. And on the runtime side, it maintains continuous behavioral baselines for sessions, tool calls, and destinations, layers detection with identity, data sensitivity, posture, permission, and exposure signals, and rounds things out with attack-path reconstruction, data-impact visibility, and remediation guidance.