Owners removed the secret from 14 of nearly 1,700 public Claude Code settings files in the week after Apiiro's report about these files. Another 7 files went offline with their repositories. All the others are still on GitHub, and the API keys, tokens and server passwords are still in them.

The files come from a Claude Code habit that Apiiro described on November 5. When you approve a command with "Yes, and don't ask again", Claude Code saves that exact command to .claude/settings.local.json. If the command had a database password or an API key in it, the file now holds it too. That file often gets committed with the rest of the project. Apiiro counted about 1,100 public copies of it with a real secret. The post reached the Hacker News front page on November 7. I took my first snapshot about 20 hours later.

What changed in a week

I checked every file three more times over the next week.

Files with secrets (1,667) Nov 9 Nov 11 Nov 15
File deleted 2 6 12
Secret removed, file kept 0 1 2
File edited, secret still there 0 3 5
Repository deleted or made private 1 3 7

The cleanups came at a steady pace. Half of them happened by November 11, and the other half by November 15. To tell a reaction from ordinary churn, I also followed about 2,500 settings files that hold no secrets.

In the week after the report With secrets Without secrets
File deleted 0.7% 0.2%
File got any commit 0.9% 2.7%
Repository deleted or made private 0.4% 0.7%

Owners deleted files with secrets about four times as often, so the report did reach some of them. Repositories with secrets went private or disappeared no more often than the rest, so I can't count those 7 as a response. Files with secrets also got a third as many commits as other settings files. I think many of these repositories have no active owner who'd read a security report.

Deleting the file leaves the secret in git history

I could trace most of the cleaned files to a commit made after the report. In every one of them, the older commit that contains the secret is still in the branch history. Anyone who clones the repository gets it. If you find a secret in your own settings file, Apiiro's advice is to rotate it first and clean up the file second. From the outside I can't tell whether these owners rotated anything.

Adding .claude to your .gitignore doesn't help once the file is committed. One in five of the files with secrets were in repositories whose root .gitignore already listed .claude. Git still tracks them, because ignore rules only apply to files that aren't in the repository yet. You can see what Git tracks in that folder with git ls-files .claude. To stop tracking a file, you need git rm --cached on it, and the old commits still keep the secret.

Elsewhere on GitHub the reaction peaked within three days

Outside my sample, the report clearly reached people. I counted commits whose message names settings.local.json with GitHub commit search. In the two weeks before the report, there were about 340 a day. In the first three days after the Hacker News post, there were about 570 a day. Over the next three days the count was still about 470.

On ordinary days, 11 to 18% of these messages mention secrets, keys, tokens or leaks. In the first two days after the post, the share was 41%. It fell to 32% on the third day. After that it stayed between 18 and 26%. In my sample, only 15 files with secrets got a commit during the whole week.

How I collected the data

I ran 54 GitHub code searches. Each one paired a term that usually appears next to a secret with settings.local.json or settings.json under .claude. The 27 terms included sk-ant, ghp, AKIA, Bearer, PGPASSWORD and sshpass. GitHub caps a query at 1,000 results, so I split the large ones by file size. That returned about 21,700 files.

A script read each file at the commit the search had indexed. It flagged values shaped like a provider API key (604 files), an access token (803 files) or a password for a remote host (260 files). Placeholders, references to environment variables, passwords for localhost, expired tokens and the public Supabase demo keys didn't count.

I didn't store, test or publish any secret. For each value the script kept its type, its length and a salted hash. That was enough to tell later whether the same value was still in the file. It also means I can't say whether any of these secrets still work.

This is my own sample, not Apiiro's list. Code search only sees default branches, and its index updates within hours, so files cleaned before my first snapshot never made it into the sample. The files without secrets are a fixed slice of the search results, the ones whose repository and path hash to one value out of seven. The commit message shares come from the first 300 messages GitHub returns for each day.