Elastic found attackers using Claude Code and the ChatGPT desktop app as shells in live intrusions to steal credentials
Elastic's quarterly 'threats, etc' report, covering confirmed live intrusions between 4 April and 8 September 2026, found attackers running commands through trusted AI agents. On 5 August 2026, Elastic confirmed an operator driving Claude Code through its VS Code extension to stage tooling and tunnel to a suspicious GitLab instance. In another case on a Mac, the ChatGPT desktop app was used as a shell command proxy to read an internal infrastructure document and dump the login keychain. Two more macOS cases used permission-bypass flags, Cloudflare tunnels and persistent LaunchAgents; Elastic's endpoint detection caught the activity.
What the AI did
Attackers who had already got into victims' machines used Claude Code, an AI coding assistant, and the ChatGPT desktop app to run their commands for them. Through these tools they staged attack software, opened hidden network tunnels, read an internal infrastructure document and dumped a Mac's stored login passwords.
First reported November 5, 2026 · Added to the register November 11, 2026 · 1 source
- Model
- ChatGPT desktop app, Claude Code
- When it happened
- August 5, 2026
- First reported
- November 5, 2026
What this means for you
Could this affect you?
If you allow AI coding agents such as Claude Code or the ChatGPT desktop app on staff computers. Organisations that have approved AI agents on their computers are exposed, because once attackers are inside they can use these trusted, vendor-signed tools to run commands, steal credentials and keep tunnels open while looking like normal employee activity.
What to check
- Inventory the AI agents approved on your computers.
- Give each AI agent a unique identity you can revoke.
- Monitor what AI agents actually do rather than trusting them because they are signed or approved.
- Alert on permission-bypass flags, tunnelling and password-store access started by AI agent processes.
Areas of your AI programme this touches
Timeline
- August 5, 2026Happened
- November 5, 2026First reported
Every fact and its source (6)
- Claude Code intrusion confirmed2026-08-05“On 5 August 2026, Elastic threat hunters confirmed an operator on a Windows endpoint”[1]
- Permission bypass flag used--allow-dangerously-skip-permissions“in Claude Code's case, --allow-dangerously-skip-permissions”[1]
- DetectionElastic Defend rules fired on claude.exe child processes“Elastic Defend rules fired on the PowerShell children spawned by claude.exe”[1]
- Report coverage period4 April to 8 September 2026“between 4 April and 8 September 2026”[1]
- Credential theft actionRead internal QA document and dumped login keychain“One compound command read lines 20 to 45 of an internal QA infrastructure document, then dumped the login keychain”[1]
- Tunnel tool usedSOCKS5 proxy script minsk-tunnel.ps1“a SOCKS5 proxy tunnel script named minsk-tunnel.ps1”[1]
Sources
How this record is classified. Severity S (4/5): real harm to one organisation, data exposed, or a containment failure. OECD level: incident, an event in which an AI system led to actual harm. Evidence: Confirmed, meaning a company, official or court statement.
The facts in At a glance and in Every fact and its source are quoted from the sources listed. The summary, What the AI did and What this means for you are written from those sources and checked against them automatically before publication. Records reflect what has been disclosed, not everything that has happened. Incidents are found through news feeds and the GDELT Project.

