Rethinking Security for Agentic AI

Rethinking Security for Agentic AI argues that as software can act autonomously or semi-autonomously, security must move from static policy enforcement to real-time behavioural governance. Agentic AI introduces new threats beyond prompt misuse, including action-level exploits, context injection attacks, invisible operations, and protocol vulnerabilities linked to open settings in models and MCP‑related interactions.

The piece notes high-profile examples such as a wiper‑style prompt injection affecting an Amazon coding assistant, and concerns around exploits like EchoLeak and CurXecute that enable access to internal data, external communication, and inputs from untrusted sources. To mitigate these risks, organisations should implement guardrails that trace actions, verify intent, and apply contextual controls across platforms, rather than focusing on a single LLM.

It also recommends building a taxonomy of agents by initiation, deployment, connectivity, and autonomy to prioritise controls, and adopting a hybrid security approach that combines deterministic policies with real‑time observability. A four‑pillar security framework is proposed: discovery and profiling, agentic posture management, observability, and runtime controls.

Written by Etay Maor, the article emphasises that enterprises must rethink risk, threat models, and protections to manage autonomous software effectively.