Researchers have found that a 'don't ask again' setting in Anthropic's Claude Code coding assistant can save approved commands, including any passwords they contain, in plain text inside a project file, and that many such files have already been exposed publicly on GitHub.
Security · AI coding tools
A "don't ask again" setting can leave passwords in plain text in project files
Approved commands saved by Claude Code can include credentials without masking. Many of the resulting files have already been exposed on public GitHub repositories.
1,122
Publicly accessible files on GitHub containing secrets
379
Database passwords found in those copies
23
AWS keys that can grant access to cloud accounts
Secrets found in exposed copies, to scale
Column height is proportional to the count of each secret type.
Takeaway: database passwords outnumber AWS keys roughly 16 to 1 in the exposed copies, but each AWS key can reach an entire cloud account.
How a convenience setting becomes public exposure
STEP 1
User approves a command
→
STEP 2
"Don't ask again" saves it unmasked
→
STEP 3
File is committed to a shared repo
→
STEP 4
Public copies get copied and scanned by others
Why it matters
Two common habits combine into one risk: routine approvals and shared configuration.
Repeatedly approving routine commands to avoid interruptions
Storing project configuration in shared code repositories
Exposed credentials can be copied long before the owner notices
What developers should do
Audit project directories and public repositories for stored credentials
Rotate any keys that may have been exposed
Remove affected files from repository history
Avoid saving approved commands that contain passwords
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...