Vulnerability / research

Apiiro finds credentials in published Claude Code settings and agent transcripts

Reported

Published here

What happened

Apiiro found 1,122 public copies of a Claude Code project settings file that contained credentials under its detection criteria. The file can preserve sensitive values typed into commands approved for future use.

Parties and reporting

AI tools and systems: Claude Code permission settings, Public GitHub repositories and their history, Saved coding-agent transcripts

Outcome

Across local and shared settings, file histories and agent transcripts, Apiiro identified 2,539 distinct apparent secrets in 1,994 repositories. It notified organization owners, one of whom confirmed its key had been live.

Operating lesson

Rotate exposed credentials before removing them from files and history. Keep secrets out of saved commands and include agent settings in repository secret scanning.

Evidence and limits

The researchers inspected public material November 1–3. They classified values by format or context and report that they never tested credentials or contacted target hosts. No malicious use or complete remediation count was established.

Sources