Cybersecurity has always been a contest between defenders and attackers operating with similar tooling. AI breaks that symmetry — and not always in the defender's favour. Understanding the new equilibrium is now a board-level responsibility.
Defenders gain leverage
AI agents triage alerts, summarise incidents, generate detections from threat intelligence, and reduce mean time to respond by orders of magnitude. The shortage of skilled security analysts has been the binding constraint on most security operations centres. Agents partially relax that constraint.
Attackers gain industrialisation
The same capabilities that help defenders also help attackers. Spear-phishing at scale, polymorphic malware, vulnerability research, and social engineering are all materially cheaper. This does not change the fundamental defender's playbook — it raises the floor of attacker capability and lowers the cost of attempts.
Agents themselves become assets
An autonomous agent with credentials, tool access, and decision authority is a new class of identity that must be managed: scoped permissions, action audit logs, kill switches, anomaly detection on agent behaviour, and clear ownership. Most enterprises have nothing comparable to identity governance for agents. That is an exposed flank.
New attack surfaces
Prompt injection, data exfiltration via context windows, model supply-chain compromise, and tool misuse are all real and exploitable. Security architectures must evolve from network-and-identity to network-and-identity-and-model.
Practical priorities for CISOs
Inventory every agent and AI integration in production. Apply identity governance to agents with the same rigour as humans. Red-team agents the way you red-team networks. Require evals and guardrails before any agent goes to production. Rehearse incident response for AI-specific failure modes.
Investment implications
AI security is becoming its own category alongside identity, network, and data security. The fastest growing segments are agent governance, model security, AI red-teaming, and runtime monitoring. Each has a clear buyer and a clear pain point.
Key takeaways
- AI raises both defender and attacker capability — net effect is asymmetric.
- Agents are a new identity class and must be governed accordingly.
- Prompt injection and tool misuse are real, exploitable attack surfaces.
- AI security is emerging as a distinct, fast-growing investment category.