
AI is evolving rapidly, especially in the enterprise, and looks very different from how it did just a few years ago. Early generative AI tools mostly answered questions or produced content in response to a prompt. Today’s AI agents, however, can act independently, using software, accessing business systems, and carrying out tasks on a user’s behalf.
The fact that interactions are no longer contained within a chatbot changes the security question. Humans and AI agents are working across the same tools and systems, often with access to sensitive information or existing user permissions. This creates a new AI interaction layer between humans, AI agents, tools, and systems, where activity must be understood as it happens.
AI Detection and Response (AIDR) is emerging as a way to secure that layer, extending detection and response into AI systems as they become increasingly autonomous.
The security challenge around AI used to be relatively contained. Much of the focus was on protecting prompts and stopping sensitive data from being exposed. Security teams also had to think about what a model might return to the user and whether that response broke policy.
Agentic AI changes that considerably. Tools such as Claude Code and OpenClaw can run directly on endpoints and do far more than generate a response, such as executing code, accessing files and using the permissions of the person who deployed them. That means an AI interaction can now lead to something happening elsewhere in the organisation, often at machine speed.
A useful parallel is endpoint detection and response (EDR). As endpoint activity became harder to secure with preventative controls alone, EDR gave security teams visibility into what was happening on the device and a way to respond when behaviour became suspicious. Agentic AI is creating a similar need for visibility at runtime, particularly on the endpoint where many of these agents now execute.
Most existing security controls weren't designed for AI systems that can act across an organisation’s environment. Traditional controls can still protect data or restrict access, but they don't always show how an agent behaves once it starts using tools, making decisions, or interacting with other systems.
For Australian organisations, that gap matters as AI adoption often moves faster than the ability to govern it. CrowdStrike cites research showing that 45% of employees use AI tools without informing their manager or employer. Even where AI governance policies exist, 61% of organisations say they lack the technology to enforce them. The threat itself is also becoming more tangible, with nearly 16,000 confirmed AI-related security incidents recorded in 2025.
The challenge is not simply whether an AI tool has access to sensitive information. Security teams also need to understand what the agent is doing with that access and whether its
behaviour changes over time. Existing tools can provide part of that picture, but they were not built to monitor the full AI interaction layer or respond to agent activity in context.
AI Detection and Response (AIDR) is designed to give security teams visibility into AI activity as it happens, so they can detect suspicious behaviour and respond before it spreads. The emphasis is on what an AI system is doing, rather than relying only on controls applied before or after the event.
That means looking beyond the model itself. CrowdStrike’s AIDR architecture covers seven layers of the AI estate, including the data AI relies on and the models being used. It also covers prompts and agents, the identities involved, and the infrastructure and interactions around them. The idea is to connect activity across those layers rather than treating each one as a separate security problem.
AIDR then applies that protection across the places where AI operates. CrowdStrike groups these into three coverage planes: SaaS, where AI connects with business data; the endpoint, where agents execute; and cloud environments, where AI workloads are built and scaled.
Running across those layers is the AI Control Plane, which is intended to turn visibility into action. Depending on the risk, that can mean enforcing access, protecting sensitive data or stopping activity at runtime. CrowdStrike also describes capabilities that support development, governance and threat detection across the AI estate.
This is where the comparison with EDR becomes particularly useful: AIDR takes the same basic idea of observing behaviour and responding in real time, but extends it across the wider AI environment.
AI security is moving beyond prompts and outputs. As agents gain more autonomy and begin acting across endpoints and cloud environments, organisations need to understand what those agents are doing in real time, not just what they were asked to do.
That is the gap AIDR is intended to address. For security teams, the immediate priority is to understand where AI is already being used across the organisation, including activity that may sit outside formal approval processes. From there, they can assess whether existing controls provide enough runtime visibility and how agent identities and permissions are governed. That should also make it easier to identify behaviour that could create new exposure before it becomes an incident.
If you want to explore AIDR in more detail, CrowdStrike’s AIDR: Defining the Next Era of Cybersecurity sets out the model and the thinking behind it. Meanwhile, if you’re reviewing what agentic AI means for your own security requirements, Infotrust can help you assess the gaps and strengthen your approach.