Understanding the Security Risks of AI Agents: Exposing the New Attack Surfaces

The emergence of AI agents has significantly altered the security landscape, exposing multiple attack surfaces that traditional models do not account for. Standard prompt attacks are only the beginning, as these agents not only generate text but also plan, use tools, and maintain memory across sessions. A 2026 report indicates that 98 percent of cybersecurity leaders experience friction between the rapid adoption of agentic AI and security requirements, leading to constrained deployments. Unlike standalone LLMs, which have a single attack surface, AI agents expose four distinct surfaces, each necessitating unique threat models. Vulnerabilities arise when agents fetch data from untrusted sources, as indirect prompt injection can occur, allowing attackers to manipulate inputs that the agent treats as trusted. Effective defense measures must focus on securing the inputs to agents and managing the parameters they can manipulate.