Agent security splits into three layers, each becoming a business of its own

2026.10.15Source · Industry watch

In October three companies moved at three different layers at once: runtime isolation, protocol gateway protection and automatic vulnerability repair. Locking down agents is no longer a feature — it is a category forming on its own.

What each layer covers

  • Runtime isolation: Perplexity open-sourced the SPACE sandbox, which shuts an agent inside an isolated virtual environment so it cannot step outside its bounds to alter the host system or third-party sites; partners including CoreWeave, Akamai and Crusoe are already deploying and reselling it.
  • Integration-layer protection: Radware launched MCP Server Protection, aimed at protocol abuse between agents and enterprise systems, folded into its existing LLM firewall and Agentic AI Protection.
  • Vulnerability layer: Cohere acquired AISLE, using AI to find and patch the software flaws an agent might reach.

Why it suddenly matters

The attack surface grows with the number of connectors. As MCP servers went from single digits to more than 1,200, every one of them is a live, authenticated channel — and traditional web and DDoS defences do not know how to look at that channel. The three layers have to coexist because no single vendor can claim to cover an agent end to end.

The human side is shifting too

In Barracuda’s October skills-gap report, 56% of organisations had cut security roles because of AI, split almost evenly between junior and senior; 69% run incident response as “AI or automation plus human investigation”. Tools are filling the gap, but judgement still sits with people.

What it means for enterprises

  • Define the list of systems an agent may reach before discussing which tool to buy. Without that list, a tool is only reassurance.
  • Log every tool invocation so it can be replayed to a specific decision — if it cannot be traced back, it is not under control.
  • Keep human confirmation for payments, deletions and public publishing; do not leave those to probability.
  • Put any agent that connects to third-party tools in a sandbox. Isolation is not an option.

Want to talk through what you just read?

We work end to end, from product design to launch. Tell us about your scenario and we can start there.

Shenzhen XZS Technology Co., Ltd. · Tel 0755-82494862 · 13760424861

Contact us

← Back to News