Engineering AI Security Starter: Stop Credentials and IP Leaking Through Your Dev Team

Developer adoption of AI coding assistants (GitHub Copilot, Cursor, ChatGPT) has outpaced security policy at most organisations. Developers paste production configs, API keys, database connection strings, and proprietary algorithms into AI tools every day — often without realising the risk.

This guide covers what to protect and how to configure Pretzel's engineering-specific policy rules.

The Threat: What Developers Actually Paste

Based on patterns observed in the Pretzel platform, the most common sensitive content in developer AI prompts falls into four categories:

1. API Keys and Tokens The single most common finding. Developers copy-paste .env files, config snippets, or Terraform files to ask AI for help — and include live production credentials in the context. AWS access keys, GitHub personal access tokens, Stripe secret keys, database passwords.

Risk: If the AI provider retains inputs (standard terms on most free tiers), your production credentials are in their training data or logs.

Pretzel rule: High-entropy detection combined with known key format patterns. Pretzel's entropy detector catches keys even when they're not in your keyword list — it looks for strings that are statistically random enough to be a credential.

2. Database Connection Strings A complete database URL contains host, port, username, and password in a single string. These appear in prompts when developers ask AI to help debug queries or schema issues.

Pretzel rule: Pattern rule for connection string formats (postgres://, mysql://, mongodb://, and similar). Action: Block.

3. Internal Project Names and Codenames Unreleased product names, M&A target codenames, and internal project identifiers can constitute material non-public information or trade secrets. Discussing an unreleased project's architecture with ChatGPT exposes the project's existence.

Pretzel rule: Keyword blocklist of internal codenames. Maintained by your security team and pushed to Pretzel. Action: Warn with justification required.

4. Proprietary Algorithms and Business Logic Novel algorithms, pricing engines, and proprietary models pasted into AI for debugging represent genuine IP risk. Some AI providers claim broad rights to use API inputs.

Pretzel rule: This is harder to detect automatically. Consider requiring acknowledgement for prompts over a certain length threshold, or scoping AI assistant access to approved tools only.

Configuring the Engineering Template in Pretzel

The Pretzel Engineering template includes: - High-entropy detection (enabled by default for the Engineering subject) - AWS, GCP, and Azure key format patterns - Database connection string patterns - A starter internal project name blocklist (empty by default — populate with your organisation's codenames)

Scope recommendation: Apply the entropy detection rules broadly to all engineering staff. Apply the proprietary codename blocklist more selectively to senior engineers and product leads with access to unreleased roadmap information.

Getting Started

Activate the Engineering template from the Pretzel Console. Add your internal project codenames to the keyword blocklist. Review the audit log after the first week to see what's being caught — and adjust sensitivity as needed.