securitycontrols.ai

Field notes

Working notes on securing autonomous AI agents — the threats, the controls, and how to prove they hold. Companion writing to the Agentic Security Control Matrix.

17 posts

Start here

The argument+1 coming

An update

The six layers

LAYER · IDENTITY

Identity and authority: who the agent is, and what it can do

The first layer of the matrix. Five controls that decide which agent did what, and what each one was allowed to do, plus how to prove each holds.

LAYER · CONTAINMENT

Environment and containment: where the agent runs, and what it can reach

The second layer of the matrix. Nine controls built on one assumption: sooner or later the agent will be wrong or hijacked, so bound the blast radius before it happens.

LAYER · PROTOCOL

Inter-agent and tool protocols: how the agent talks to tools and to other agents

The third layer of the matrix. Seven controls for the wires between an agent and everything it talks to, where the supply chain and the injection risk concentrate.

LAYER · GOVERNANCE

Governance and human-in-the-loop: who approves, and what is on the record

The fourth layer of the matrix. Nine controls for the human checkpoints, the records, and the accountability that keep an autonomous system answerable.

LAYER · RUNTIME

Runtime supervision and detection: what the agent actually does, watched live

The fifth layer of the matrix. Eight controls for watching an agent as it runs, including the frontier ones nobody has fully solved yet.

LAYER · ASSURANCE

Continuous assurance: proving it before it ships, and after

The sixth layer of the matrix. Seven controls that prove an agent is safe to run, before it goes live and re-proven on every change after, the gates, the supply-chain checks, and the proof that it holds at all.

The vendor landscape

VENDOR-LANDSCAPE · IDENTITY

How the identity layer is being built

A snapshot of who is building agent identity and authorization, organized by the controls they map to. Standards first, vendors second, no recommendations.

VENDOR-LANDSCAPE · CONTAINMENT

How the containment layer is being built

A snapshot of who is building agent containment, organized by the controls they map to. Standards first, vendors second, no recommendations, and an honest map of where no product exists yet.

VENDOR-LANDSCAPE · PROTOCOL

How the inter-agent and tool-protocol layer is being built

A snapshot of who is building the protocol layer, organized by the controls they map to. Standards first, vendors second, and an honest note on how much of the security is still missing.

VENDOR-LANDSCAPE · GOVERNANCE

How the governance layer is being built

A snapshot of who is building agent governance, organized by the controls they map to. Standards and regulation first, vendors second, no recommendations.

VENDOR-LANDSCAPE · RUNTIME

How the runtime layer is being built

A snapshot of who is building the runtime supervision layer, organized by the controls they map to. Standards first, vendors second, no recommendations.

VENDOR-LANDSCAPE · ASSURANCE

How the continuous assurance layer is being built

A snapshot of who is building continuous assurance, organized by the controls they map to. Standards and open work first, labs and vendors second, no recommendations.

VENDOR-LANDSCAPE-CAPSTONE

Convergence, new energy, and the seams

Stepping back from the six layers: where the field is consolidating, where the genuinely new energy is, and who is, and is not, moving on the three gaps no framework closes.