Model misalignment is when an AI system’s behavior diverges from its intended, approved, or safe objectives. Instead of following the operator’s policy, it may optimize for the wrong goal, obey harmful instructions, or take actions that were never meant to be allowed. In security terms, this is not just a correctness bug; it is a control problem.
Misalignment matters because modern AI agents can browse, call tools, edit content, or trigger workflows. If a model is exposed to malicious text, prompt injection, or overly broad permissions, it may act outside its scope and create real operational risk. Defenders reduce this risk with narrow privileges, action confirmation, output filtering, monitoring, and strong logs that make unexpected behavior visible quickly. In practice, misalignment can look like unsafe advice, unauthorized writes, leaked context, or an agent taking a public-facing action without proper authorization.