Your agent's permission file is a secrets ledger

Apiiro and GitGuardian found thousands of live credentials in .claude/settings.local.json on public GitHub. Why deleting the file fixes nothing, and the rotation order that does.

Your agent's permission file is a secrets ledger

A developer needs to check staging. The agent proposes PGPASSWORD=hunter2 psql -h staging-db.internal -U app -c "select count(*)", and they answer "Yes, and don't ask again" because they will be doing this all afternoon.

That click writes the entire command — password included, plain text — into .claude/settings.local.json. The file is JSON, it sits next to the code, and it is one git add . away from permanent.

Two studies have now measured what happens next.

What the studies found

Apiiro scanned public GitHub and found 104,185 copies of that file across 103,617 repositories. About 1.1% — 1,122 files — held at least one secret, 1,379 in total. More than a third open a database directly (379 database passwords), plus 43 SSH logins, 42 GitHub, GitLab and npm tokens and 23 AWS keys, across 63 organisations. One affected file held the owner password of a medical-records database. The median leaking file had sat on its default branch for 217 days.

GitGuardian ran the wider query: continuous monitoring from January 2025 to 7 November 2026 found 31,088 distinct secrets in the same file across 23,276 repositories, 5,343 still valid at detection — 17%. The live bucket is mundane: 1,955 distinct valid Postgres connection strings, 941 Google API keys, 186 DeepSeek keys.

Apiiro counted what is exposed today without testing validity; GitGuardian counted everything ever pushed. The second number is your rotation backlog.

The protection is machine-local

Claude Code is not careless by accident. The first time it writes the file inside a git repository, it adds the path to the user's global git excludes. Git on that laptop then ignores the file everywhere.

That is where it stops. The rule is not in the repository's own .gitignore, so nothing travels with the project. A folder that becomes a repo later, a fresh clone on a colleague's machine, a file committed before the rule existed — none inherit it. In 539 of the 1,105 repositories Apiiro traced, Claude was a co-author on the commit that changed the file.

Boris Cherny, who leads Claude Code, says the team is investigating and that exposed copies were likely committed by hand. GitGuardian's rebuttal: across roughly 3,000 occurrences, every secret sat in an allow rule of the exact shape the permission prompt writes.

Deleting the file fixes nothing

Deleting the file does not un-leak a single credential. Old versions sit in git history — 734 further secrets across 509 repositories — and in 211 forks of 74 repositories. Rewriting history does not reach forks, and it never reaches a container registry.

Docker is the quiet third exit. GitGuardian found 109,284 occurrences inside public Docker Hub images across 7,625 tags, 72% under /app, /workspace or /usr/src. COPY . . copies the working tree, not the index, so a file git ignores still ships unless .dockerignore lists .claude/. One valid Cloudflare token appears in 559 tags of a single image.

Why your scanner stayed quiet

Generic scanners look for shapes: AKIA prefixes, PEM blocks, JWTs, high entropy. On the 1,122 leaking files, gitleaks flagged 564 and TruffleHog 643, and 215 were flagged by neither. A password after PGPASSWORD= has no provider prefix; it is an ordinary word in an ordinary command line.

GitGuardian's placement data maps onto that gap. curl is half of all occurrences — mostly bearer tokens, low validity. The credentials that still work cluster where format scanners read worst: psql, mysql and mongosh rules valid more than half the time, sshpass -p at 54%.

The gateway key is the expensive one

.claude/settings.json is the file teams commit on purpose. A quarter of its leaking versions carry an env block pointing Claude Code at a third-party Anthropic-compatible gateway — ANTHROPIC_BASE_URL plus ANTHROPIC_AUTH_TOKEN aimed at DeepSeek, OpenRouter, Z.ai, MiniMax or Kimi. The credential that leaks is the gateway's: 205 valid DeepSeek keys, 62 OpenRouter, 50 MiniMax. Real Anthropic keys are rare in both files.

That is a billing risk before it is an access risk — someone else's token bill, someone else's rate limits, and a request log you will be explaining to a client.

What we changed ourselves

We run coding agents on our own repositories daily, so this is our hygiene problem too. The reframe that stuck: a permission file is production state with a lifetime, and git ignoring it on the machine that created it is not a control — the same lesson as the agent attack surface and the canary work on AI routers.

Second, separate revocation from cleanup. Rotating a key finishes in one API call; removing a file is a promise about copies you do not control. Reverse the order and you will feel done while a fork still holds the credential. We also treat .claude/, .cursor/, .codex/, .mcp.json and CLAUDE.md as secret-bearing: GitGuardian found 6,155 distinct secrets in CLAUDE.md alone, 2,537 valid.

The checklist

  1. Search every repo for .claude/settings.local.json and .claude/settings.json, including deleted files in history.
  2. Rotate every credential those files ever held — before touching git.
  3. git rm --cached .claude/settings.local.json, then add it to the repo's own .gitignore.
  4. Add .claude/ to .dockerignore; check images already in the registry.
  5. Delete every local rule containing a secret — /permissions lists them.
  6. Reference secrets as variables ($DATABASE_URL), so the rule records the name, not the value.
  7. Approve narrow patterns (Bash(npm run:*)), never Bash(curl:*), which green-lights any curl a prompt injection writes.
  8. Teach scanners rule grammar: a VAR=value prefix, a password flag or userinfo in a URL is a finding at any entropy.
  9. Audit settings.json env blocks for gateway tokens and cap their spend.
  10. Prefer auto mode with sandboxing over permanent approvals — Anthropic's guidance, and what 90% agent-written code requires.

None of this argues against the tools. It is the pattern from every agent finding this year: permission state, transcripts, memory files and handoff notes were convenience features first, and they are now where credentials live.

Header photo: Rawpixel, CC0 1.0.