Elastic reports that intruders used Claude Code on Windows and ChatGPT desktop on macOS to carry out activity through applications their organizations already trusted. Its November 4 report announcement introduces the full November threat report. The report describes two August cases, rather than two newly occurring November breaches.

Evidence status: customer intrusions confirmed by Elastic’s own threat hunters and reported through its telemetry. The public report does not independently establish a worldwide attack rate, a flaw in either application’s signing process, or how attackers initially obtained access.

Approved applications supplied an execution path

On August 5, Elastic says an operator used Claude Code’s VS Code extension over a remote desktop connection on a Windows endpoint. The activity staged tools and contacted a suspicious self-hosted GitLab instance through a proxy tunnel. Elastic’s endpoint rules detected shell processes and network clients beneath the Claude process.

On August 10, threat hunters identified credential access through the ChatGPT desktop application on macOS. A shell sequence read internal infrastructure documents and attempted a Keychain dump filtered toward that infrastructure. The Keychain is macOS’s store for passwords and other secrets.

These are examples of misuse after access, not evidence that installing an AI assistant automatically lets an outside attacker enter a workstation. Applications with permission to run tools inherit access to files, credentials and network destinations. A legitimate signature identifies the application; it does not authorize everything an operator asks it to do.

The report does not identify affected or fixed application versions, a product CVE, or a repair release for these cases. Updating software remains normal maintenance, but no specific upgrade can be presented as this report’s complete fix.

A tunnel alert still needs investigation

Elastic’s related August research on coding-agent tunnels and LaunchAgents is more cautious about a different macOS sequence. It reconstructs credentialed requests, a reverse tunnel exposing a local application, and LaunchAgents that kept services running across sessions.

That article explicitly leaves two possible readings: unauthorized access, or legitimate remote administration of a local dashboard. The tools and destinations are dual-use. A LaunchAgent is a macOS mechanism for starting user processes, so persistence through one is not itself proof of malware.

Keep that uncertainty separate from the November report’s two confirmed August intrusion findings. A tunnel-provider reputation label, a shell under an agent or a suspicious-looking command cannot settle intent by itself.

Investigate the outcome, not just the parent name

Elastic recommends correlating process ancestry with credential access, network destinations and persistent changes. For defenders and developers, the practical checks are:

  • Keep sensitive outcomes visible. An approved coding agent should not automatically suppress alerts about credential collection, unexpected tunnels or persistent startup changes.
  • Review the whole session. Compare the activity with the authorized task, actual destinations and the user’s explanation. Separate routine tool execution from unexplained access to secrets.
  • Limit the agent’s reachable authority. Map which files, accounts and services its execution environment can access. Grant the access needed for the task, with deliberate approval for broader changes.
  • Respond to established exposure. If investigation confirms unauthorized access, preserve evidence and contain it through the normal response process. Revoke or rotate credentials actually exposed; removing the application cannot recall copied secrets.

Those are controls on authority and response, not a claim that every tunnel or agent command is an intrusion.

Caveat: Elastic sells endpoint protection and supplies the underlying findings; the public report omits customer identities and raw telemetry needed for independent verification.