Security

This guidance helps you use AI tools securely in government software development.

Key security risks when using AI

How to check AI-generated code for vulnerabilities

AI-generated code can contain security vulnerabilities or accidentally embed secrets like API keys.

Do:

• Review code for vulnerabilities, especially those found in the OWASP Top 10 (opens in new tab)
• Check code for any embedded secrets like API keys or database credentials
• Scan generated code with a Static Application Security Testing (SAST) tool
• Ensure that security-specific code like encryption is carefully reviewed

Don’t:

• Assume AI-generated code is free of vulnerabilities
• Trust that AI-generated code implements security-critical functionality correctly

How to secure dependencies in AI-generated code

AI-generated code can import packages which have known vulnerabilities or which are planted by attackers.

Do:

• Scan all dependencies with a Software Composition Analysis (SCA) tool to highlight any vulnerabilities
• Use constraints such as “Use secure and up-to-date dependencies” in your prompts
• Check any unfamiliar packages for trustworthiness using a service like deps.dev (opens in new tab)

Don’t:

• Assume dependencies included in AI-generated code are secure or non-malicious

How to protect your development environment

Local AI tools like Cursor can expose you to risk in certain circumstances.

Do:

• Prevent local user credentials and auth tokens from being exposed to the AI tool (for example with a .cursorignore file)
• Only install AI tool plugins maintained by trusted organisations
• Use only approved MCP servers when connecting AI tools to external systems—MCP is currently in review and limited to pilot projects (see Defra MCP Guidance)

Don’t:

• Use any untrusted data when writing prompts or user stories that end up in prompts
• Install untrusted plugins
• Auto-approve MCP tool calls without review

Next -> Getting Started