CrowdStrike Adds AIDR Controls for Copilot Studio and Claude Code

For Copilot Studio agents using generative orchestration, AIDR can act as an external threat-detection provider. Before the agent executes a tool, Copilot Studio sends invocation context to AIDR. CrowdStrike says AIDR evaluates the tool name and input parameters against an assigned policy, then returns an allow or block decision.

The product documentation narrows that scope. AIDR does not analyze the user prompt or recent conversation history even though Copilot Studio includes them in the request, and it does not inspect tool outputs. The external interface supports only whole-invocation allow or block decisions; transformation rules operate in report-only mode. Copilot Studio also imposes a one-second response limit, after which the configured error behavior determines whether the invocation is allowed or blocked. Classic agents with fixed action flows are not supported.

The Claude Code collector integrates with the product's hook system. It can evaluate a prompt before Claude processes it, inspect a tool call before execution, and evaluate tool output after completion. CrowdStrike documents allow, block, or transform decisions for the supported hook events, with blocking available at prompt submission and before or after tool use.

Deployment requires adding an HTTP-hook configuration to Claude Code settings rather than installing a separate binary. Administrators can place the configuration in managed settings to make the policy apply across users on a machine.

The Claude Code path is also bounded. It does not monitor model responses, model configuration, token usage, extended thinking, system prompts, or subagent context. It can redact tool inputs and outputs but not submitted user prompts. CrowdStrike says the resulting activity can be correlated with endpoint, identity, and network telemetry in Falcon Next-Gen SIEM.

For security teams, the release adds enforceable checkpoints around two widely used agent surfaces, but it is not complete agent observability. Policy design should account for the fields each integration cannot see, the Copilot timeout behavior, and the difference between blocking a tool action and inspecting the reasoning or output that led to it. CrowdStrike's announcement and documentation describe product behavior; they do not provide independent detection-accuracy or latency results.