Every AI agent is a security identity

Agentic AI changes cybersecurity on both sides.

Traditional AI answers questions and provides recommendations. AI agents go further. They receive goals, make plans, access data, invoke tools, communicate with other agents, and execute actions.

This creates three cybersecurity priorities:

• Protect AI agents
• Defend against AI-enabled attacks
• Deploy AI agents within security operations

The threat is already accelerating.

CrowdStrike reported an 89% year-over-year increase in attacks involving AI-enabled adversaries.

At the same time, Microsoft used an agentic security system to identify 16 vulnerabilities across the Windows networking and authentication stack, including four critical remote-code-execution flaws.

Agentic AI strengthens attackers and defenders.

The difference will come from execution.

Every enterprise agent needs:

• A unique identity
• Least-privilege access
• Short-lived credentials
• Approved tools and data sources
• Defined financial and operational limits
• Continuous behavioral monitoring
• Complete action logs
• Human approval for high-impact decisions
• Immediate containment and shutdown controls

Cybersecurity vendors are moving quickly:

• Cisco acquired Robust Intelligence (now part of Cisco) and integrated its technology into Cisco AI Defense.
• Palo Alto Networks acquired Protect AI, Koi, and Portkey.
• CrowdStrike acquired Pangea and introduced Falcon AI Detection and Response.
• Fortinet announced the acquisition of Virtue AI.
• Check Point Software introduced its AI Defense Plane and AI Agent Security solution.
• Microsoft introduced Project Perception, an agentic security system using specialized agents to reason across security data and workflows while preserving human control over critical decisions.
• Google introduced specialized agents in Google Security Operations for threat hunting, detection engineering, third-party risk analysis, investigation, and response.

The market direction is clear. AI security is expanding beyond protecting prompts and models.

The new control plane must cover:

• Agent identities
• Permissions and credentials
• Tools and MCP servers
• Data access
• Inter-agent communication
• Runtime behavior
• Endpoints and cloud infrastructure
• Security operations

A compromised chatbot might expose information. A compromised agent might access systems, change configurations, move data, initiate payments, or instruct other agents to act.

Security teams also need agents to investigate alerts, correlate intelligence, identify vulnerabilities, recommend responses, and execute approved containment actions at machine speed.

Humans will remain responsible for strategy, exceptions, high-impact decisions, and accountability.

Every AI agent is a new workforce member. Every workforce member needs an identity, permissions, supervision, and accountability.

Is your security architecture ready to govern thousands of AI agents?