What new categories of privacy/security risk do modern AI (LLM) systems introduce?

Prompt injection & jailbreaks, hallucinations, data leakage via context, agent escalation & tool abuse, and brand/reputational harm.

Modern AI systems now operate beyond the user interface, interacting directly with tools, data, and APIs — language has become both the interface and the control layer. As these systems gain agency, they bring new risk categories:

• Prompt injection & jailbreaks — hijacking the model's instructions.
• Hallucinations & confabulation — confident false outputs.
• Data leakage via context — secrets in the context window exposed.
• Agent escalation & tool abuse — invoking APIs/plugins with attacker-controlled parameters.
• Brand & reputational harm — the model says something damaging.

The underlying problem: organizations deploy AI faster than they can verify its trustworthiness.

Tip: Where there's privacy break-points: training-data leaks, malicious prompts, harmful outputs, verbose system logs, and untrusted third-party plugins.