Navigating the Agentic AI Security Landscape

Generative AI has rapidly shifted from interactive assistants to autonomous, goal-directed agents that plan, act, delegate to sub-agents, and persist over time. In enterprise environments these agents can write and execute code, call APIs, manage files, and orchestrate workflows, creating a qualitative expansion of the risk surface beyond traditional model-inference concerns. While input filtering, output moderation, and prompt-injection defenses remain necessary, they are insufficient because agentic deployments introduce a second security surface centered on action, identity, and delegation.

A six-category framework organizes the agentic AI security and governance landscape across layers from identity and access through runtime control to governance and compliance, mapped to four operational phases: Prevent (AIAM + APGE), Observe (AVOP), Detect/Respond (ATDR), and Govern (ARC + AORC). Procurement also splits into three distinct tracks with different buyers: Track 1 extends Generative AI Defense to the agent layer; Track 2 addresses AI Agent Identity & Access Management; Track 3 focuses on AI Governance, Risk & Compliance, driven by EU AI Act requirements beginning August 2026 for high-risk systems.

Agent identity creates new challenges: agents and sub-agents are short-lived, hierarchically organized, and often invisible to existing IAM/IGA. Trust anchors and attestable workload identity (not long-lived secrets) become central, with SPIFFE/SPIRE and cloud workload identity options highlighted. Authorization is an active frontier: preventing privilege amplification across delegation chains requires multi-tier policy-based access control, with OpenFGA and OPA positioned as enabling policy engines though full multi-tier policy propagation is not yet implemented by vendors. Market dynamics show rapid consolidation in ATDR (comparable to CASB and SOAR waves), pushing buyers to evaluate integrated platforms rather than standalone ATDR products and to prioritize credential-sprawl audits and EU AI Act-driven auditability and risk documentation.

• AI agents in enterprises can write and execute code, call external APIs, manage files, and orchestrate workflows.
• The document distinguishes four terms: Generative AI, AI agent, Agentic AI, and avoids the unqualified term “agent.”
• Controls designed for GenAI inference are necessary but not sufficient for agentic deployments.
• The six categories map to four operational phases: Prevent (AIAM + APGE), Observe (AVOP), Detect/Respond (ATDR), and Govern (ARC + AORC).
• The Generative AI Defense Leadership Compass was published in December 2025 and evaluates LLM and agent runtime protection.
• The vendor examples listed are illustrative as of April 2026 and are not endorsements.
• Between September 2024 and Q4 2025, seven vendors across the landscape were acquired, including five in ATDR and adjacent APGE categories.
• EU AI Act obligations for high-risk AI systems in the EU begin in August 2026, including Article 9 risk management, Article 10 data governance, and Article 13 transparency requirements.
• SPIFFE/SPIRE is presented as a leading open standard for issuing short-lived workload identities.
• SentinelOne and CrowdStrike (via Pangea) are identified as major platform vendors without a fully integrated AI agent security acquisition.

• “The shift from generation to agency is not incremental.”
• “It is a qualitative change in the risk surface that enterprises must manage.”
• “The policy question is no longer only ‘is this request permitted?’ but ‘what is this agent permitted to do, with what identity, on what data, across what delegation chain, with what accountability trail?’”
• “The categories are not sequential (enterprises may have gaps in any combination), but they follow a logical architecture from identity and access through runtime control to governance and compliance.”
• “Several vendors appear across multiple categories.”
• “ATDR generates the signal; the AI SOC is the operational layer that consumes and acts on it.”
• “Agentic AI exposes a richer identity problem than traditional non-human identity (NHI) governance has previously addressed.”
• “No vendor has fully implemented multi-tier policy propagation; this is the most active research frontier in the AIAM category.”
• “The immediate practical exposure is credential sprawl.”
• “Buyers should assess platforms with integrated ATDR capabilities, focusing on completeness of functionality, depth of integration with the wider security stack, and continuity of the acquired technology…”