threatdeck
Live 13:45 UTC ·

Detection library

KQLHunting query: tune before alertingAI coding agents on endpoints

Destructive cloud or cluster CLI command launched under an AI coding agent

A malicious prompt shipped in Amazon Q Developer for VS Code 1.84.0 (AWS-2025-015) told the agent, run with --trust-all-tools --no-interactive, to wipe the home directory and use AWS profiles to list and delete cloud resources; it failed only because of a syntax error. This hunt finds aws, az, gcloud, kubectl and terraform commands that destroy resources when the parent or grandparent is an AI agent.

Log source
Microsoft Defender for Endpoint: DeviceProcessEvents
Techniques
AML.T0101 Data Destruction via AI Agent Tool Invocation · ASI02 Tool Misuse and Exploitation · LLM06 Excessive Agency · ASI10 Rogue Agents
MITRE ATT&CK
T1485Data DestructionT1578.003Delete Cloud InstanceT1059.009Cloud API
Reviewed
2026-11-10

Query

let ai_agents = dynamic(["claude", "claude.exe", "codex", "codex.exe", "cursor-agent", "Cursor", "Cursor.exe", "Cursor Helper (Plugin)", "gemini", "copilot", "Windsurf", "Windsurf.exe", "q"]);
let ai_package_regex = @"(@anthropic-ai[\\/]claude-code|@google[\\/]gemini-cli|@openai[\\/]codex)";
DeviceProcessEvents
| where Timestamp > ago(30d)
| where InitiatingProcessFileName in~ (ai_agents) or InitiatingProcessParentFileName in~ (ai_agents)
     or InitiatingProcessCommandLine matches regex ai_package_regex
     or InitiatingProcessFolderPath contains "/.local/share/claude/versions/"
| where ProcessCommandLine matches regex @"(?i)(ec2\s+terminate-instances|s3\s+rb\b|s3\s+rm\b.*--recursive|iam\s+delete-(user|role|access-key)|rds\s+delete-db-(instance|cluster)|cloudformation\s+delete-stack|az\s+group\s+delete|gcloud\s+projects\s+delete|kubectl\s+delete\s+(ns|namespace)\b|terraform\s+destroy)"
| project Timestamp, DeviceName, AccountName, InitiatingProcessParentFileName, InitiatingProcessFileName,
          FileName, ProcessCommandLine
| order by Timestamp desc

Expected false positives

  • Engineers asking an agent to tear down their own sandbox or test stack
  • Agents run inside ephemeral environments whose credentials only reach throwaway accounts

Sources

Reviewed against vendor documentation, not tested in your environment. Defensive use only.

Following

Follow the vendors, products, AI tools, techniques and actors you care about. Matching stories, CVEs and exploited bugs are marked, a Following tab appears in the feed, and returning visits show what changed. Saved only in this browser.

Presets

Get updates

A personal RSS feed of new stories on everything you follow. Paste it into any feed reader, Slack or Teams RSS app.