Every SaaS application your organization runs may be individually secure, compliant, and audited. It doesn't matter. Security does not compose cleanly. Anthropic's Mythos model proved it — and the proof is permanent.
In the 1990s, anthropologist Robin Dunbar observed that humans can maintain stable relationships with roughly 150 people at once. Beyond that number, the cognitive load becomes unmanageable. We lose track. We lose context. We lose the ability to reason about the whole.
The same limit applies to complex systems.
A CISO can hold one application's architecture in their head with clarity. Maybe two. By the time you're reasoning about five or ten interconnected enterprise platforms simultaneously — their APIs, their data flows, their trust relationships, their shared identity layers — you are at the edge of what human cognition can manage. So organizations did what humans always do when they hit a cognitive limit: they partitioned.
Each silo felt like a boundary. Each team felt like a perimeter. The organization felt secure because every domain had an owner and every owner had a handle on their domain.
We call this the Dunbar Perimeter — the implicit security boundary created not by architecture, but by the cognitive limit of the humans responsible for it.
The modern enterprise SaaS ecosystem didn't create the Dunbar Perimeter — it industrialized it.
Every SaaS vendor sells a best-of-breed solution for a specific domain. CRM. ERP. HRIS. ITSM. Finance. Procurement. Collaboration. Each one is optimized, supported, and secured within its own boundary. Each one has its own compliance certifications, its own security team, its own penetration testing program.
And each one connects to all the others.
The average enterprise now runs hundreds of SaaS applications. Each integration — each API call, each SSO connection, each shared data flow — creates an edge between two nodes. That edge crosses a Dunbar Perimeter. It belongs to neither team. It is secured by no single owner. It lives in the gap between two cognitive domains that were never designed to reason about each other.
This was always true. The reason it wasn't catastrophic was simple: the attacker had a Dunbar limit too. Human attackers, even sophisticated ones, could only hold so many systems in their heads simultaneously. They picked targets. They worked domains. The cognitive partitioning that limited defenders also limited attackers.
That symmetry is gone.
Anthropic's Mythos model does not have a Dunbar limit. It does not partition systems into cognitive domains. It does not fatigue. It does not privilege one context over another.
It constructs and traverses the full dependency graph — across identity systems, SaaS platforms, internal services, and data flows — simultaneously, autonomously, and without the cognitive overhead that made siloed security feel adequate.
The connections that were previously invisible to defenders are now first-class attack paths. Not new connections. The ones that were always there — running through every integration boundary, every API edge, every cross-system workflow that was owned by nobody because it crossed a Dunbar Perimeter.
The inversion is complete:
You can no longer evaluate Salesforce's attack surface independently from SAP's. They are not separate surfaces. They are nodes in a single traversable graph. They always were. The difference is that now, something can see and exploit that graph in its entirety — and it works overnight while your team sleeps.
The Dunbar Perimeter is gone. Hardening individual applications inside individual Dunbar silos is no longer a viable strategy. The response is not more controls inside each silo. It is elimination of the silos — and the edges between them.
Consolidate, Reduce, Close. Three directives. In order. Each one removes material that a Mythos-class adversary requires to traverse your stack.
A CRC assessment applies exclusively to governed server deployments under formal organizational control. End-user devices, developer workstations, and any environment where AI agents with computer-use capability operate outside a governed boundary are outside CRC scope. If the Governed Server Boundary prerequisite is not met, the MSS is not zero -- it is inapplicable.
Not hardening. Elimination. Four pillars. Total score 0 to 16. Regulated deployment certification requires 13 or higher. The era of the Dunbar Perimeter is over. Any architecture that depended on human cognitive limits to feel secure is now exposed — because the adversary no longer shares those limits.
CRC is published as an open standard. Score your own stack. Read the framework. The goal is not General Reasoning's business — it is a regulated infrastructure posture that can actually survive the adversary that now exists.