A rogue AI agent is an automated AI workflow that acts beyond its intended permissions or guardrails. This may happen because access controls are too broad, instructions or integrations are misconfigured, or safeguards fail to constrain the agent’s actions. The term describes behavior, not necessarily a self-directed or conscious system.
In cybersecurity, an agent with access to email, files, code repositories, or business systems can expose data or make unauthorized changes if its actions are not properly limited. Attackers may also manipulate an agent through malicious inputs or compromised connected services. Defenses include least-privilege access, approval steps for sensitive actions, limits on what tools and data an agent can reach, and logging that lets teams review its activity.