Defensive Mesh: Control When Attackers Become Agents

SOCHQ White Paper

Architecture & Finished Platform

Final Draft · 2026-08-13

Abstract

Machine-speed, adaptive offensive agents collapse the time from probe to impact. Defensive systems that separate sensing from action, or that export critical context and reasoning outside the customer boundary, cannot keep pace and cannot satisfy enterprise requirements for residency, controllability, and independent verification.

A viable architecture must keep continuous sensing, specialized always-on agents, coordinated orchestration, and execution capability on the same assets. It must operate as a state-determined system: the next defensive action is determined by the current measurable state of the estate, the active campaign objects, and the governing policy—not by hidden operator memory, chat history, or unconstrained agent improvisation. Governance must enable speed rather than obstruct it; unconstrained action must be earned by evidence and remain revocable.

SOCHQ is built to these requirements. It is a defensive mesh platform whose control plane enrolls and governs an on-estate workforce—RMM hands, XDR sense, stack stewards, and AI co-workers. The system notices attack paths, expands dens on those paths, interrupts under authority, and proves every step with client-verifiable audit. The architecture is a closed-loop, state-determined response to the joint demands of tempo and control.

1. The Shift

Defensive operations assumed a human-scale tempo for two decades. That assumption is obsolete.

Offensive capability now includes adaptive, multi-vector agentic systems that maintain state, learn from defensive responses, and pursue multiple paths in parallel. The interval between foothold and impact is compressing into the same timescale on which the attacker operates.

When the offensive loop runs at machine tempo, any architecture whose critical path depends on exporting telemetry, waiting for centralized investigation, or routing every consequential decision through an open-ended human process is already late. Visibility is not response. Recommendation is not containment. The requirement is continuous sensing, rapid formation of a coherent response from the current state of the estate, and execution on the assets themselves—under residency, authority, and audit constraints that enterprises will accept.

This is a systems problem. The correct response is a state-determined defensive system: next action is a function of measurable state, active path objects, and explicit policy, closed by feedback, and capable of interrupting under authority without improvising outside its model.

2. Structural Misalignment of Current Models

Cloud-centric agentic SOCs place the decisive reasoning step outside the customer boundary. They introduce latency, create a residency and trust surface at the moment of highest sensitivity, and make independent verification difficult.

Pure endpoint agent swarms can act locally and quickly, but global context remains weak, policy consistency is hard to enforce, and unconstrained local autonomy is rarely acceptable.

Hybrid models keep sensors on the assets yet still export the critical investigation step. Local presence is retained while the decisive cognitive step is externalized.

None of these patterns keeps continuous sensing, specialized review, coordinated orchestration, and execution capability together inside a governed, state-determined boundary. That combination is required.

3. Non-Negotiable Requirements

• Continuous sensing and specialized review must run on the assets.
• A coordination layer must receive continuous signal and dispatch targeted work without becoming the sole source of truth.
• Execution capability must operate on the same assets that generate sensing and review.
• Primary context remains under governed residency, with an on-estate inference path available when policy requires it.
• The system must be state-determined: next actions follow from current estate state, campaign objects, and policy.
• Unconstrained (“free-act”) authority is earned by evidence on a defined action class and remains revocable. Default is dual-run or approval.
• The full loop—from notice through interrupt—must produce a client-verifiable, hash-chained audit surface.

Against adaptive, machine-tempo offensive agents, any defensive architecture that cannot keep continuous sensing, specialized review, coordinated orchestration, and execution capability together on the estate—under an explicit authority model and client-verifiable proof—will be limited to visibility or delayed response. The SOCHQ mesh is built to close that loop.

4. SOCHQ: Control Plane + Defensive Mesh

SOCHQ is a defensive mesh platform designed as a state-determined control system for the estate.

A control plane enrolls and governs an on-estate workforce—RMM hands, XDR sensors, stack stewards, and specialized AI co-workers. Bodies phone home under signed policy. They share estate identity and campaign context through the plane. The plane maintains the dual graph (estate graph + work graph); chat sessions are ephemeral, the graph persists. Inference is brokered under policy—cloud or on-estate private mesh—so that tempo-critical and residency-sensitive judgment can remain inside the trust boundary.

Foundation Mesh

• **XDR Sensor** — process, network, DNS, file, registry, and auth dens; causal lineage; Layer-1 always-on notice plus Layer-2 dens-pull on path; interrupt surface.
• **RMM Agent** — universal hands: inventory, health, signed task execution, deploy, contain/isolate, service and package control under policy. The agent verifies task signature before every act.
• **Stack Stewards** — continuous health, propose, gated act, and exception-native recovery on Wazuh, Graylog, and stack-class peers.

Constant Fabric

Always-on reviewers operate across the mesh and the plane: estate bodies, coverage and readiness watches, House Analyst, L1 SOC, NOC, Ticket Analyst, and Stewards. A Red Team agent supplies continuous authorized pressure findings into the platform knowledge base as fuel.

Control Plane

Orchestration is a core function of the SOCHQ Control Plane. It maintains tenant-scoped master context, receives signals from the constant fabric, and dispatches dens expand, specialist work, and response verbs. It is the system of record for control, judgment, and proof—not a second SIEM lake.

The Defensive Loop

Notice → Continuous Review → Path Object / Expand → Propose → Gate → Execute (host · process · network · identity) → Receipt + Re-measure.

5. Five Defensive Planes

1. Continuous Adversarial Pressure — Red Team agent feeds path priors into the platform knowledge base.

2. Estate Readiness & Blast Radius — Stewards, RMM, and enrollment keep sensors live and kill-paths present.

3. Notice + Path Expand — XDR two-layer sensing plus L1 campaign recognition.

4. Interrupt Under Authority — Host, process, network, and identity act verbs under the free-act matrix, dual-run, and approval ladder.

5. Proof & Recovery — Receipt, re-measure, evidence package, and client/auditor export.

State is continuously updated. Next action is determined by that state under policy.

6. Human Gate and Autonomy Model

Free-act is a promotion, never a default. This is product law.

The human gate is required for governance and proof. It is also the mechanism that generates the dual-run evidence needed to promote free-act class by class. Autonomy is earned with measured evidence rather than assumed.

High-blast false positives are treated as first-class system failures. The architecture defaults to dual-run and approval for irreversible or high-blast verbs. Promotion is earned only on a measured safe matrix (tenant × verb × blast radius × scorecard) and is automatically demoted on regression. Every free-act decision is fully receipted and re-measured.

7. Observability, Accountability, Traceability, Auditability

The Control Plane is the system of record for control, judgment, and proof.

Every material action leaves a complete lineage: who authorized, what ran, what changed, what was verified. The lineage is recorded in a per-tenant hash-chained audit log. The customer can recompute the chain; a single-byte tamper fails at the exact position. Decision trajectories are exportable. The design principle is absolute: “Don’t trust us — check us.”

We concentrate control so that authority remains coherent and auditable. We then engineer the plane and the bodies so that concentration does not become a single point of catastrophic failure: bodies safe-idle when the plane is unreachable, fleet and tenant emergency stop, dual venues, on-estate brain tier, short-lived credentials, signed task verification on every act, and surviving audit.

8. What the Architecture Delivers

• Continuous notice and campaign recognition on the estate
• Path expand with Layer-2 dens-pull when a chain is active
• Interrupt across host, process, network, and identity under explicit, revocable authority
• On-estate brain capacity under plane policy
• Client-verifiable proof of every material step
• Typed coverage honesty—no synthetic green

The platform is a state-determined defensive control system that retains authority when attackers operate as agents.

9. Closing

The tempo of offense has changed. Architectures that treat defense as open-loop alerting, externalized reasoning, or unconstrained local autonomy cannot close the loop at the required speed while remaining acceptable to the enterprises that must run them.

The structural solution is a state-determined mesh: continuous sensing and specialized review on the assets, coordinated orchestration through a control plane, execution under an explicit free-act and approval regime, and proof the customer can independently verify. Body forward. Brain brokered. Interrupt under authority. Graph persists.

SOCHQ is built to that design. It is the coherent structural response to the joint demands of tempo and control.

Positioning Language (Final)

Offensive AI agents collapse the time from probe to path complete.

SOCHQ is the defensive control plane: enroll and govern bodies on the estate, expand on the attack path with XDR, interrupt under dual-run and free-act law, and prove every step—continuously improved by red-team intel into the platform knowledge base, with optional on-estate brain capacity under plane policy.

We concentrate control so authority remains coherent and auditable, then engineer the plane and the bodies so that concentration does not become catastrophic failure.

We are not building the ultimate attacker. We are building the system that still has control when attackers become agents.

Preferred terms: SOCHQ · Control Plane · RMM · XDR · Stack Stewards · defensive mesh · state-determined · free-act matrix · dual-run · governed residency · body forward, brain brokered · “Don’t trust us — check us”

Doctrine locks:

• Body forward, brain brokered
• Free-act is a promotion, never a default
• Dual graph — chat dies, graph stays
• State-determined
• Mesh stays on the estate
• Don’t trust us — check us