AI safety is the discipline of keeping AI systems reliable, bounded, and safe when they move from test environments into real operations. In cyber security, it matters because a model can be accurate yet still dangerous if it can access sensitive data, trigger tools, or make decisions without clear limits and review.
Defenders apply AI safety through scoped permissions, logging, human approval for high-risk actions, and ongoing testing for prompt injection, data leakage, and unsafe automation. Attackers target weak AI safety by manipulating prompts, poisoning data, or exploiting overbroad tool access to make defensive systems reveal information or take harmful actions. Good AI safety turns AI from an unchecked control surface into a monitored, reversible part of the security stack.