<p>I decide what an AI assistant needs before I provide code, logs, screenshots, or tool access. Most debugging work needs structure and error context, not a customer's message, full token, private hostname, or complete production record.</p><p>Redaction is part of preparing the task. I replace identifying values with consistent synthetic ones so relationships remain visible without exposing the original data.</p>

<h2>Classify the material</h2>

<p>I mark credentials, personal data, client-confidential data, internal infrastructure, proprietary code, and public information. The classification determines whether the material can be used, must be reduced, or should stay out of the workflow.</p><p>A value can be harmless alone and identifying in combination. A timestamp, rare route, and exact error may reveal a customer incident even after the name is removed.</p>

<h2>Redact while preserving structure</h2>

<p>I replace tokens with <code>TOKEN_REDACTED</code>, accounts with stable fictional IDs, domains with <code>example.invalid</code>, and user records with synthetic fixtures. Consistent replacements let the assistant trace one identity across a sanitized log.</p><p>I remove authorization headers and query strings before copying a request. Screenshots receive the same review as text.</p>

<h2>Use the smallest code scope</h2>

<p>I share the failing function, type, test, and necessary configuration instead of an entire proprietary repository when broad access is not required. If Codex has repository access, the task and filesystem scope still state what it may inspect and change.</p><p>External systems remain read-only during diagnosis unless the user authorizes a bounded write.</p>

<h2>Review generated commands</h2>

<p>I resolve targets before executing commands that deploy, delete, rotate, migrate, or send messages. I reject unresolved variables, broad globs, and output that may print environment values. Commands run in a non-production environment first when possible.</p><p>A model can produce syntactically correct destructive commands. Human review supplies authority and context.</p>

<h2>Respond to accidental exposure</h2>

<p>If a secret enters a prompt, log, commit, or screenshot, I revoke or rotate it. Deleting the visible copy does not prove downstream copies disappeared. I review access logs and affected systems according to the incident plan.</p><p>Client data exposure follows the organization's privacy and notification procedure. The developer should not improvise legal conclusions.</p>

<h2>Use a pre-share checklist</h2>

<p>I pause before every upload or pasted block and search for common credential formats, email addresses, hostnames, account numbers, and message content. Automated secret scanning helps but does not replace context review.</p><ul><li>Only necessary files are in scope.</li><li>Examples use synthetic values.</li><li>Tool permissions match the task.</li><li>Commands have explicit targets.</li><li>Exposed credentials are rotated.</li></ul>

<h2>Need a safer development workflow?</h2>

<p>Terry Hunt Dev can help teams define safer automation and development practices through <a href="/services/devops-consulting">DevOps consulting</a>. Discuss the policy through the <a href="/contact">contact page</a> without pasting private logs or credentials.</p>

Improve AI coding assistant security with rules for redacting secrets and client data, using synthetic examples, limiting access, and reviewing commands.

Keeping Secrets and Client Data Out of AI Coding Workflows

Redact credentials and customer data, use synthetic examples, restrict tool scope, and review commands before AI-assisted work changes systems.

AI Security Resources & Benchmarks

Keeping Secrets and Client Data Out of AI Coding Workflows

Need hands-on help with AI Security?