Back to Intelligence

Securing AI Workflows: Deploying Falcon AIDR for Copilot Studio and Claude Code

SA
Security Arsenal Team
July 31, 2026
6 min read

The integration of Generative AI into the enterprise workflow has shifted from experimentation to critical infrastructure. With the release of Falcon AIDR (AI/LLM Detection and Response) now covering Microsoft Copilot Studio Agents and Anthropic’s Claude Code, CrowdStrike has closed a critical visibility gap. As these tools gain autonomy to read codebases, execute shell commands, and access internal data, they become prime targets for indirect prompt injection and jailbreaking. Defenders can no longer treat AI traffic as benign web traffic; we must inspect the context and intent of LLM interactions in real-time.

Technical Analysis

The Attack Surface Unlike traditional web applications, AI agents like Copilot Studio and Claude Code possess "tool use" capabilities. An attacker does not need to exploit a buffer overflow; they simply need to manipulate the LLM's context to trigger authorized tools for malicious purposes (e.g., exfiltrating data via a Claude Code shell command or manipulating business logic via a Copilot agent).

Affected Platforms & Components

  • Microsoft Copilot Studio: Low-code/No-code agents interacting with Dataverse and external APIs. Vulnerable to indirect injection via untrusted data sources ingested by the agent.
  • Anthropic Claude Code: A CLI/IDE-integrated agent capable of executing terminal commands and manipulating file systems. Vulnerable to jailbreaks leading to arbitrary code execution (ACE) via the AI interface.

The Threat Vector Falcon AIDR addresses LLM Input/Output Inspection. The security challenge lies in detecting "polymorphic" prompt attacks where the malicious intent is obfuscated. The Falcon Sensor intercepts these interactions to identify:

  1. Indirect Prompt Injection: Malicious content hosted on websites or documents read by the AI.
  2. Jailbreaking: Bypassing safety guardrails to force restricted actions.
  3. Data Exfiltration: Attempts to pipe sensitive corporate data out of the LLM context window.

Detection & Response

While Falcon AIDR provides the automated control plane for these threats, security teams must maintain visibility into the usage of these AI tools to ensure policy compliance and detect potential tool abuse before the agent is fully compromised. Below are detection mechanisms to monitor the deployment and execution of Claude Code and anomalous agent behaviors.

SIGMA Rules

YAML
---
title: Potential Claude Code Jailbreak via Shell Execution
id: 8a4c1d22-5f6e-4b8a-9c1d-2e3f5a6b7c8d
status: experimental
description: Detects Claude Code or associated IDE processes spawning unauthorized shell commands, a common indicator of successful prompt injection or tool misuse.
references:
 - https://www.crowdstrike.com/en-us/blog/falcon-aidr-protects-copilot-studio-agents-and-claude-code/
author: Security Arsenal
date: 2026/04/06
tags:
 - attack.execution
 - attack.t1059.001
 - attack.t1204
logsource:
  category: process_creation
  product: windows
detection:
  selection:
    ParentImage|contains:
      - 'claude.exe'
      - 'Code.exe' # VS Code host for Claude extension
    Image|endswith:
      - '\cmd.exe'
      - '\powershell.exe'
      - '\pwsh.exe'
  filter_legit_dev:
    CommandLine|contains:
      - 'git'
      - 'npm'
      - 'build'
  condition: selection and not filter_legit_dev
falsepositives:
  - Legitimate developer build processes
level: high
---
title: Suspicious Network Traffic to Anthropic API from Non-Browser
id: 9b5d2e33-6g7f-5c9b-0d2e-3f4g6a7b8c9d
status: experimental
description: Identifies unexpected processes communicating with Anthropic API endpoints, which may indicate unauthorized usage of Claude Code or data exfiltration attempts.
references:
 - https://www.crowdstrike.com/en-us/blog/falcon-aidr-protects-copilot-studio-agents-and-claude-code/
author: Security Arsenal
date: 2026/04/06
tags:
 - attack.command_and_control
 - attack.t1071.001
logsource:
  category: network_connection
  product: windows
detection:
  selection:
    DestinationHostname|contains:
      - 'api.anthropic.com'
    Initiated: true
  filter_browsers:
    Image|endswith:
      - '\chrome.exe'
      - '\msedge.exe'
      - '\firefox.exe'
  filter_claude_code:
    Image|endswith:
      - '\claude.exe'
  condition: selection and not filter_browsers and not filter_claude_code
falsepositives:
  - Custom scripts integrating with Anthropic API
level: medium

KQL (Microsoft Sentinel / Defender)

This query hunts for high-volume data transfers or specific error codes associated with AI agent interactions that might indicate a successful data dump or repeated jailbreak attempts.

KQL — Microsoft Sentinel / Defender
let AnthropicEndpoints = dynamic(["api.anthropic.com", "claude.ai"]);
let CopilotEndpoints = dynamic(["copilotstudio.microsoft.com", "copilot.microsoft.com"]);
DeviceNetworkEvents
| where RemoteUrl in (AnthropicEndpoints) or RemoteUrl in (CopilotEndpoints)
| where Initiated == true
| summarize SentBytes = sum(SentBytes), ReceivedBytes = sum(ReceivedBytes), ConnectionCount = count() by DeviceName, InitiatingProcessFileName, RemoteUrl
| where SentBytes > 5000000 or ConnectionCount > 100 // Thresholds for data exfil or automated abuse
| project DeviceName, InitiatingProcessFileName, RemoteUrl, SentBytes, ConnectionCount, Timestamp
| order by SentBytes desc

Velociraptor VQL

Hunt for the presence of Claude Code binaries and inspect the configuration for potential policy violations or unauthorized installations.

VQL — Velociraptor
-- Hunt for Claude Code and AI Agent Binaries
SELECT FullPath, Size, Mtime, Mode
FROM glob(globs="/*/*/claude.exe")
WHERE Mode =~ "r-x"

-- Cross-reference with running processes
SELECT Pid, Name, Exe, Username, Ctime
FROM pslist()
WHERE Name =~ "claude" OR Exe =~ "claude"

Remediation Script (PowerShell)

This script verifies that the CrowdStrike Falcon Sensor is running and at a version compatible with AIDR features, and audits the local environment for the presence of Claude Code.

PowerShell
# Audit and Remediation Check for AI Agent Environment
Write-Host "[+] Starting AI Agent Security Audit..." -ForegroundColor Cyan

# 1. Check CrowdStrike Falcon Sensor Status (Required for AIDR)
$csService = Get-Service -Name "CSFalconService" -ErrorAction SilentlyContinue
if ($csService.Status -eq 'Running') {
    Write-Host "[+] CrowdStrike Falcon Sensor is Running." -ForegroundColor Green
} else {
    Write-Host "[!] ALERT: CrowdStrike Falcon Sensor is not running. AIDR protection is inactive." -ForegroundColor Red
}

# 2. Audit for Claude Code Installation
$claudePaths = @(
    "$env:LOCALAPPDATA\Programs\Claude",
    "$env:USERPROFILE\AppData\Local\Programs\Claude"
)

$foundClaude = $false
foreach ($path in $claudePaths) {
    if (Test-Path $path) {
        Write-Host "[!] Found Claude Code installation at: $path" -ForegroundColor Yellow
        $foundClaude = $true
    }
}

if (-not $foundClaude) {
    Write-Host "[+] No Claude Code installations found in standard user paths." -ForegroundColor Green
}

# 3. Check for Restricted API Keys in Environment Variables (Generic Security Check)
$envVars = Get-ChildItem Env:
$apiKeyPatterns = @("ANTHROPIC_API_KEY", "OPENAI_API_KEY")
foreach ($pattern in $apiKeyPatterns) {
    if ($envVars.Name -contains $pattern) {
        Write-Host "[!] WARNING: API Key pattern '$pattern' found in environment variables." -ForegroundColor Red
    }
}

Write-Host "[+] Audit Complete." -ForegroundColor Cyan

Remediation

To fully mitigate the risks associated with AI agents and utilize the new Falcon AIDR capabilities:

  1. Update Falcon Sensors: Ensure all endpoints are running the latest CrowdStrike Falcon Sensor version that supports the AIDR module for Copilot Studio and Claude Code.
  2. Enable AI-Streaming Policies: In the Falcon console, navigate to Falcon AIDR > Settings and enable inspection for "Copilot Studio" and "Claude Code". Ensure traffic is being proxied or intercepted correctly by the sensor.
  3. Configure "Allow" vs "Audit" Modes: Initially, run AIDR in Audit mode to understand baseline AI usage. Switch to Block mode once legitimate prompts are baselined to stop jailbreaks automatically.
  4. Review API Key Hygiene: Ensure that Claude Code instances are not hardcoding service account credentials with excessive privileges (e.g., Global Admin). Use least-privilege keys for AI tool integrations.

Related Resources

Security Arsenal Penetration Testing Services AlertMonitor Platform Book a SOC Assessment vulnerability-management Intel Hub

Is your security operations ready?

Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.