Friendly Fire: Mitigating Autonomous AI Agent Code Execution Risks

---
title: Potential AI Agent Autonomous Code Execution
id: 8a2c1d05-6e9f-4c3b-9a1e-2f4c5d6e7f8a
status: experimental
description: Detects AI coding agents spawning shell processes to execute scripts in temp directories, indicative of "Friendly Fire" auto-execution behavior.
references:
  - https://www.hackernews.com/2026/07/friendly-fire-ai-agents-built-to-catch.html
author: Security Arsenal
date: 2026/07/16
tags:
  - attack.execution
  - attack.t1059.004
logsource:
  category: process_creation
  product: linux
detection:
  selection_ai_parent:
    ParentImage|endswith:
      - '/python'
      - '/node'
      - '/claude'
      - '/codex'
    ParentCommandLine|contains:
      - 'claude'
      - 'openai'
      - 'agent'
  selection_shell_spawn:
    Image|endswith:
      - '/bash'
      - '/sh'
      - '/python'
    CommandLine|contains:
      - '/tmp/'
      - '/var/folders/'
      - 'T/'
  condition: all of selection_*
falsepositives:
  - Legitimate developer testing of scripts in temp directories
level: high
---
title: AI Agent Spawning Suspicious Child Process (Windows)
id: 9b3d2e16-7f0a-5d4c-0b2f-3g5d6e7f8g9b
status: experimental
description: Detects Windows AI agent wrappers spawning PowerShell or CMD to execute scripts from AppData or Temp folders.
references:
  - https://www.hackernews.com/2026/07/friendly-fire-ai-agents-built-to-catch.html
author: Security Arsenal
date: 2026/07/16
tags:
  - attack.execution
  - attack.t1059.001
logsource:
  category: process_creation
  product: windows
detection:
  selection_parent:
    ParentImage|endswith:
      - '\python.exe'
      - '\node.exe'
      - '\Code.exe'
    ParentCommandLine|contains:
      - 'claude'
      - 'openai'
  selection_child:
    Image|endswith:
      - '\powershell.exe'
      - '\cmd.exe'
    CommandLine|contains:
      - '\AppData\Local\Temp'
      - ' -ExecutionPolicy Bypass'
  condition: all of selection_*
falsepositives:
  - Developer running local tests via AI assistance
level: high

// Hunt for AI agents spawning unauthorized shells
DeviceProcessEvents
| where Timestamp >= ago(1d)
| where InitiatingProcessFileName has_any ("python", "node", "claude", "codex") 
// Focus on common AI wrapper binaries
| where ProcessFileName in ("bash", "sh", "powershell", "cmd", "python")
| where ProcessCommandLine has_any ("/tmp/", "AppData\\Local\\Temp", "T/")
| project Timestamp, DeviceName, AccountName, InitiatingProcessCommandLine, ProcessCommandLine, ProcessFileName
| order by Timestamp desc

-- Hunt for suspicious AI agent process chains executing code in temp dirs
SELECT Pid, Name, CommandLine, Exe, Username, CreateTime
FROM pslist()
WHERE Parent.Name =~ "(python|node|claude|codex)"
  AND Name =~ "(bash|sh|powershell|cmd|python)"
  AND CommandLine =~ "(/tmp/|AppData.*Temp|T/)"

#!/bin/bash
# Remediation/Harden Script for AI Agent Environments
# This script checks for running AI agents and verifies they are not operating in potentially unsafe paths or modes.

echo "[*] Scanning for active AI Agent processes..."

# Define suspicious agent process names
AGENTS=("claude" "codex" "ai-agent")

FOUND=0

for agent in "${AGENTS[@]}"; do
  PIDS=$(pgrep -f "$agent")
  if [ -n "$PIDS" ]; then
    echo "[!] Detected running AI agent: $agent (PIDs: $PIDS)"
    FOUND=1
    
    # Check if these processes are spawning shells (basic heuristic)
    for pid in $PIDS; do
      children=$(pgrep -P $pid)
      if [ -n "$children" ]; then
        echo "    [WARNING] Process $pid has active child processes. Verify manual approval is enabled."
      fi
    done
  fi
done

if [ $FOUND -eq 0 ]; then
  echo "[+] No active AI agents detected matching current threat profile."
else
  echo ""
  echo "[REMEDIATION STEPS]"
  echo "1. IMMEDIATELY disable 'Autonomous Mode' or 'Auto-Approve' in the Agent UI."
  echo "2. Terminate the agent instances if they are reviewing untrusted code:"
  echo "   pkill -f 'claude|codex'"
  echo "3. Restart agents inside a sandboxed container (e.g., Docker) with restricted network access."
fi