Prompt Injection Basics
How attackers manipulate LLMs and how to defend against it.
Learning objectives
- Identify sensitive data that should not be entered into public AI tools
- Understand approved vs unapproved AI usage
- Recognize risky prompts
- Follow company AI policy
- Report unsafe AI usage
Lesson
What is sensitive data?
Personally identifiable information (PII), customer records, contracts, source code, financial data, trade secrets, regulated data (PHI, PCI), and any non-public business information. If you would not post it publicly, do not paste it into a public LLM.
Examples of unsafe AI usage
Pasting a signed customer contract into ChatGPT to 'summarize the key terms'. Asking an AI to debug code that contains hardcoded API keys. Uploading internal financials to generate a forecast.
Safe prompting practices
Strip identifiers, use placeholder values, and prefer enterprise-tier tools with zero-retention agreements. When summarizing documents, redact first or use an internal AI assistant.
Approved tools and workflows
Use ChatGPT Enterprise, Microsoft Copilot, Claude for Work, or the internal AI assistant. Avoid personal accounts. SSO is mandatory. Audit logs are reviewed weekly.
What to do if data was accidentally shared
Report immediately via the #ai-incidents channel or security@acme.com. Do not attempt to 'delete' the conversation — the data may already be retained. Speed of disclosure determines containment.