AI coding agent committed an API key or token into git or printed it in logs. How do I block secrets before commit with gitleaks or trufflehog, and what do I do about one that already leaked?

Cause: Agents read .env files and config, then copy values into code, test fixtures, commit messages or log output. Nothing stops a commit unless a scanner runs first. Fix (best first): 1. Add a pre-commit hook so a commit with a detected secret fails. gitleaks documents using the pre-commit framework: in .pre-commit-config.yaml add the repo https://github.com/gitleaks/gitleaks with a pinned rev (use a current release tag; pre-commit autoupdate refreshes it) and the hook id gitleaks, then run pre-commit install. A gitleaks-docker hook id runs the official image instead of a local binary. 2. trufflehog documents several hook options in its PreCommit.md. One is a hook that runs trufflehog git file://. --since-commit HEAD --results=verified,unknown --fail --trust-local-git-config, or the same with TRUFFLEHOG_PRE_COMMIT=1 set for automatic configuration. Via the pre-commit framework it can be a local hook; for Node projects it can run from Husky. The --fail flag makes trufflehog exit with code 183 when results are found. 3. Also scan in CI as a backstop, since local hooks can be skipped. gitleaks supports baselines: create a report with gitleaks git --report-path gitleaks-report.json, then scan with --baseline-path so only new findings show. 4. Keep the agent from reading secrets in the first place: deny reads of .env files in the agent's permissions (Claude Code supports deny rules such as Read(./.env)), and use a secrets manager or environment injection instead of files in the repo. 5. If a secret was already committed or logged, treat it as exposed: revoke and rotate it at the provider first. Then remove it from history only if needed (rewriting history does not undo exposure for clones, forks and caches) and review provider logs for use. Don't: use SKIP=gitleaks or git commit --no-verify as a routine workaround; the gitleaks docs mention SKIP for a single deliberate bypass only. Verify: stage a file containing an obviously fake test token pattern in a scratch repo; the commit should be rejected by the hook. Sources: https://github.com/gitleaks/gitleaks https://github.com/trufflesecurity/trufflehog/blob/main/PreCommit.md https://code.claude.com/docs/en/settings