Security teams often ask AI models the same probing questions as attackers to diagnose vulnerabilities. This triggers safety refusals, preventing them from effectively responding to incidents unless they can bypass these guardrails, as seen in the OpenAI Hugging Face breach.
The entire cybersecurity industry was built to defend against two threats: malicious people and malware. Agentic AI processes behave differently from both, representing a new category of threat that traditional signatures and behavioral analysis are not designed to handle, rendering them obsolete.
Modern deception tactics, like placing fake AWS keys (honeypots) on developer machines, are failing. An AI agent, fulfilling a legitimate user request, might innocently find and use these fake keys, triggering a massive number of false alarms and rendering the technique uselessly noisy.
Security evolved from static signatures to dynamic behavioral analysis based on the assumption that 'normal' software behavior could be defined. Agentic AI invalidates this assumption because its actions are inherently unpredictable, making it impossible to establish a reliable baseline for anomaly detection.
The current security landscape presents a paradox. While AI creates a new, complex threat surface, it also provides defenders with unprecedented tools. For example, building a software taxonomy, a task that once took years and hundreds of researchers, can now be done in weeks using AI agents.
While focus is on securing large AI models, the bigger risk is the rapid integration of agentic features into the 6,000-7,000 apps already in an enterprise. With 50% of apps projected to be agentic soon, defenders face a massive, poorly understood attack surface with no visibility into the underlying models or guardrails.