A security audit reveals that an agent tasked with generating configuration files has committed API keys, database connection strings with passwords, and private key material directly into a public repository. The agent had no secret detection configured and treated configuration values the same as any other file content. What guardrail was missing and what is the remediation?

Secret detection is a mandatory security guardrail for any agent that generates or modifies configuration files, environment files, or any content that might include credentials. Without it, an agent that is correctly following its instructions (generate a working configuration file) will produce files containing secrets, because it has no mechanism to distinguish between values that are safe to commit and values that must never enter version control.

The implementation is a pre-commit or pre-push secret scanning gate: (1) the agent prepares its commit; (2) before the commit is sent to the git remote, a secret scanner (e.g., GitHub's secret scanning, TruffleHog, GitLeaks, or detect-secrets) scans the staged content for patterns matching known secret formats (API key patterns, private key headers, connection string patterns); (3) if secrets are detected, the commit is rejected and the agent is instructed to replace the detected values with environment variable references, template placeholders, or a secrets manager call.

If secrets have already been committed, immediate remediation requires: rotating all exposed credentials (the only safe assumption is that exposed credentials are compromised), removing the commit from history (via git filter-branch or git filter-repo, then force-push with team coordination), and notifying the security team.

Option A (set repo to private) reduces exposure but does not eliminate it — private repositories can be accessed by anyone with organization access, and the credential history persists. The commit must be removed from history; privacy alone is insufficient.

Option C (lower autonomy / human review) adds a human checkpoint but humans routinely miss credentials embedded in large configuration files. Automated scanning is more reliable at pattern-matching for credential formats.

Option D (fine-tuned model to avoid generating secrets) is not a reliable safety mechanism. Models can generate secrets unexpectedly — as interpolated values, as test data, or as copied examples. Scanning is the verification layer that catches what generation-time avoidance misses.

Secret scanning as a pre-commit gate is the non-negotiable guardrail for AI-generated configuration content.