Rethinking Security for the Machine-to-Machine Internet

The Machine-to-Machine Internet

More than 50 percent of internet traffic today is automated. Not humans browsing. Machines talking to machines. APIs calling APIs. AI agents triggering workflows without a person in the loop. Most enterprise security architectures, including many built only a few years ago, were designed for the other half of that traffic, the human half.

This shift was the focus of a recent invite-only executive roundtable hosted by Cloudflare in London, which BTLITC attended alongside enterprise security leaders discussing how to secure what Cloudflare and others now call the machine-to-machine internet.

The Number That Reframes Everything

The most striking figure from the session was the change in median attacker lateral movement time. In 2022, once an attacker gained an initial foothold, it typically took more than eight hours to move laterally through a network. By 2025, that figure had fallen to roughly 22 seconds.

That collapse is not the result of more skilled attackers. It is the result of automation on both sides. When the request, the credential check and the lateral move all happen at machine speed, a security operations centre cannot review and respond in time to stop it. The enforcement has to be automated too, and it has to sit at the network layer, before the request ever reaches your infrastructure.

Where the Real Threat Lives

The conventional image of a cyber attack, someone breaking through a firewall, is no longer the primary threat model for most organisations. The more common pattern now is a valid credential on a trusted integration doing something it should not.

The Drift breach, which affected more than 700 organisations last year, made this concrete. There was no exploit and no phishing involved. It came down to one valid OAuth token on a SaaS integration that nobody was actively monitoring. The attacker did not break in. They logged in, using credentials that were never revoked or scoped tightly enough to limit the damage.

This is the essence of non-human identity risk. Every API key, service account, CI/CD pipeline and AI agent connected to your systems is effectively a digital employee with access and authority, and in most organisations, no clear owner and no lifecycle management.

What Machine Identity Governance Actually Requires

Securing the machine-to-machine internet requires the same discipline organisations have spent the last decade applying to human identity, applied to non-human identity instead. In practice that means a full inventory of every API key, service account and integration with access to your environment, ownership assigned to each one, scoped permissions that grant the minimum access required, scheduled credential rotation rather than keys that live unrotated for years, and continuous monitoring that flags unusual behaviour from an integration rather than waiting for a human-style alert.

Where MCP Fits In

The Model Context Protocol, or MCP, has accelerated this conversation considerably. MCP gives AI agents a standardised way to connect to enterprise systems, internal services, and SaaS platforms such as Slack, Jira, GitHub and Google Workspace. That is a significant productivity gain, but every new MCP connection is also a new machine identity with its own access profile.

Cloudflare's approach, discussed in detail at the roundtable, centres on MCP Server Portals: a single governed, authenticated and audited layer sitting between AI agents, MCP clients and the internal services they connect to. Rather than every agent and every client negotiating direct access to internal systems individually, traffic passes through one portal where API security, identity and data loss prevention are enforced consistently. The agent simply does its job, while the security and governance layer sits underneath, out of the way until it is needed.

This pattern matters regardless of which platform an organisation chooses. The principle is the structural one: centralise the governance layer for machine-to-machine and AI agent traffic rather than allowing every integration to manage its own access independently.

A Practical Starting Point for UK Organisations

Most UK organisations do not need to solve this problem all at once. A practical starting point is to build a complete inventory of every API key, service account and AI agent connection currently active in your environment, since most organisations are surprised by how many they find. From there, assign an owner to each one, retire anything unused, and bring the rest under a scoped access policy.

The next step is to consolidate machine-to-machine and AI agent traffic through a single governed access layer wherever practical, rather than allowing point-to-point integrations to proliferate. Finally, build credential rotation and anomaly monitoring into standard operating procedure for non-human identities, with the same rigour applied to privileged human accounts.

How BTLITC Helps

BTLITC's cybersecurity practice helps UK organisations bring machine identity and AI agent access under control, from initial discovery and inventory through to designing a governed access architecture and embedding ongoing monitoring. This work sits alongside our wider zero trust and penetration testing services, and increasingly alongside AI Vault deployments, where governing how AI agents connect to internal systems is built in from day one rather than retrofitted later. If machine-to-machine security, non-human identity governance, or platform consolidation is on your 2026 agenda, get in touch with the BTLITC team.

Frequently Asked Questions

What is non-human identity in cybersecurity?

A non-human identity is any credential, key, or token used by a machine, application, API, or AI agent to authenticate and access systems. Examples include API keys, service accounts, OAuth tokens, and CI/CD pipeline credentials. Like human identities, non-human identities require lifecycle management, scoped permissions, and monitoring to prevent unauthorised access.

How does the Model Context Protocol (MCP) affect enterprise security?

MCP gives AI agents a standardised way to connect to enterprise systems and SaaS platforms such as Slack, Jira, and GitHub. Each MCP connection creates a new machine identity with its own access profile. Without a centralised governance layer, these connections can proliferate independently, creating visibility gaps and increasing the attack surface for credential-based threats.

Why has attacker lateral movement time dropped so sharply?

Lateral movement time fell from more than eight hours in 2022 to roughly 22 seconds by 2025, because attackers now use the same automation that defenders rely on. Once an initial foothold is established, automated tooling enumerates credentials, identifies connected systems, and moves laterally at machine speed, faster than a human security operations team can detect and respond without automated enforcement at the network layer.