Check Point CCMS cheat sheet

AI Security

• AI security covers two directions: protecting AI systems (models, prompts, data, MCP/tool access) and using AI inside security products (detection, triage, policy assistance).

• Prompt injection is untrusted content in a prompt or retrieved document that redirects the model; defense is to treat retrieved content as data, never as instructions.

• Indirect prompt injection arrives via documents, web pages, or tool output rather than the user; input filtering on the user prompt alone does not stop it.

• Sensitive data leakage risk: prompts and context may contain PII or secrets sent to an external model; controls are DLP on prompts, approved-model allowlists, and redaction.

• Shadow AI = unsanctioned use of external AI services; detection relies on web/URL categorization, DNS, and SaaS discovery, not on endpoint antivirus signatures.

• Agentic AI risk: tool-calling agents can act (send mail, run code, query databases); least-privilege tool scopes and human approval for destructive actions are the core controls.

• Model output must be treated as untrusted: validate or sandbox any code or command the model proposes before execution.

• AI in security products accelerates triage and correlation but does not remove the need for human approval on high-impact actions such as blocking or isolating an asset.