The guidance has arrived
In September 2026 the Australian Signals Directorate published Agentic AI Harnesses: The layer above the model. It is one of the first government publications to name, as an architectural category, the software layer that connects an AI model to data, tools and systems — and to state plainly that this layer, not the model, is where the security of agentic AI is decided.
Its findings are direct:
“Organisations control the harness, not the LLM.” The model is a replaceable component. The layer above it is where organisational value, governance and investment accumulate, and it will outlast successive generations of models.
“No harness is inherently secure.” Established security practice — least-privilege access, identity and access management, monitoring, human oversight — must be applied to the harness as it would be to any other system connected to enterprise data.
Prompt injection has no reliable fix inside the model. Because a model cannot dependably distinguish authorised instructions from instructions hidden in the content it processes, ASD concludes that mitigations “must instead be applied in the harness, by controlling what an agent can access and what actions it is permitted to perform.”
“Do not rely on model safety controls instead of harness-enforced controls.” Enforcement must be independent of the intelligence being governed.
Record what matters. Prompts, responses, tool invocations, approvals, actions and configuration changes should be captured, protected, retained and independently monitored, with retention appropriate to risk, policy, law and incident response. The guidance also treats consumption as a security signal: unexpected cost can indicate misuse, runaway agent loops or denial-of-wallet attacks.
For readers of this research path, none of this is new territory. It is the argument of The Control Plane Must Be Yours, made by the Australian Government’s signals-intelligence and cyber-security agency.
Two and a half weeks later, on 29 September 2026, ASD followed with an alert: Protect your organisations’ AI services. It reports that malicious actors are gaining access to organisations’ AI services through compromised API keys, stolen tokens and hijacked sessions — in one cited incident, a model-provider key leaked from an internet-facing agent dashboard was used to consume roughly US$600,000 of credits in three weeks. Its framing sentence is the harness paper’s argument applied to access: “Protecting that access requires more than relying on the AI developer’s security controls.” The alert asks organisations to keep credentials out of code, documents, logs and prompts; to protect audit logs from tampering; to configure spending and consumption limits “that enforce restrictions rather than only generating alerts”; and to consider the downstream reach of a compromised agent into connected systems, tools and other agents.
One harness is never the whole estate
The guidance describes controls for a harness. Real enterprise estates run many harnesses at once.
A productivity copilot arrives with the office suite. A coding agent arrives with the development toolchain. Agent capabilities appear inside SaaS products the organisation already licenses. Delivery teams build their own. Each of these is a harness in ASD’s terms: it supplies the model’s context, executes its actions, holds its memory and enforces — or fails to enforce — its permissions.
Each harness governs itself. Each is configured separately, by different teams, on different assumptions. Each attests to its own behaviour, and some send telemetry to their vendors, depending on configuration and contract.
Applying ASD’s controls inside each harness separately therefore produces three structural problems. Policy becomes inconsistent, because it is re-implemented per product. Evidence becomes fragmented, because each harness keeps its own partial record in its own format. And verification becomes circular, because the layer being governed is also the layer producing the assurance.
Each harness governs itself. Separate products mean separate policies and separate evidence.
The controls the guidance calls for — least privilege, complete records, independent monitoring, cost limits — need a point of enforcement and evidence that does not belong to any harness or model vendor.
A direction emerging across jurisdictions
In 2026, public authorities and standards bodies converged on a practical question: how can organisations make use of increasingly autonomous AI while retaining control over the data it accesses, the systems it reaches and the consequences of its actions?
The Australian Signals Directorate and its international partners emphasise least privilege, identity, layered safeguards, human oversight, monitoring and controlled adoption. Australia’s subsequent harness guidance identifies the software connecting models to data and tools as a critical governance surface. India’s CERT-In goes further in its October advisory: evaluation environments should enforce default-deny network egress at the network layer, not through instructions to the model, while organisations should govern agent identities, destinations and activity. Singapore’s financial regulator expects firms to remain accountable for AI used in their services, including third-party AI, and to apply proportionate controls and compensating measures where appropriate.
Meanwhile, ISO/IEC 27090 provides an international reference for understanding and mitigating AI-specific cybersecurity threats. NIST in the United States is advancing work on agent interoperability, identity and authorisation. The UK NCSC continues to stress sandboxing, active oversight and planning for unexpected behaviour.
These publications differ in their legal status, audiences and technical emphasis. They do not mandate TomorrowX or a particular middleware architecture. Taken together, however, they strengthen the case for controls that an organisation can configure, operate, test and evidence outside the model’s own instructions. That is the opportunity for Data Mediation: to complement model and harness controls by applying customer-defined policy to the exchanges that are deliberately routed through it, while leaving application, network, identity and human safeguards in place. The full source library, with each document’s type and status, is on the Perspectives page.
The control plane above every harness
TomorrowX’s interpretation is that organisations should not have to repeat the same data policy separately in every harness. Where data and model exchanges can be routed through an independent Data Mediation point, they can be governed using customer-defined policies and supported by an additional source of operational evidence. This complements rather than replaces secure harness design, network enforcement, identity management and human oversight.
The pattern established at the start of this path applies directly. The enterprise places a Programmable Data Agent in the traffic path between its AI harnesses and the model services they call. The PDA terminates each interaction at the protocol level, evaluates it against enterprise policy while it is in flight, and re-originates only what is allowed — without requiring changes to the harnesses on one side or the model services on the other.
Because the enforcement point is independent, it is uniform: one policy and one evidence trail across every mediated harness and provider. Because it operates at the protocol level, it reduces dependence on the harness wherever the interfaces and routing are supported. And because it runs inside an environment the customer governs, the evidence it produces belongs to the enterprise, not to a vendor.
The guidance’s controls, enforced on the wire
Each of the practices ASD recommends has a concrete expression at an independent enforcement point in the data path.
Least privilege
Egress is deny-by-default: only approved model endpoints are reachable, and sensitive identifiers are redacted from outbound content before they leave the trust zone.
Credentials never in the application
The provider key is attached from a vault at the enforcement point, and anything the client sent in its place is stripped before the request leaves. An exposed application cannot leak a key it never held.
Records of mediated interactions
Every mediated interaction is captured at the protocol level and written to a dedicated, retention-grade audit store, separate from operational logs and independent of every harness and provider. Records are hash-chained and sealed, so any alteration is detectable to the line.
Independent enforcement
Redaction, withholding, validated identifier detection and mid-stream termination are applied in the data path, not requested of the model or the harness.
Injection mitigated above the model
Outbound content is screened against protected lists and patterns, and classification runs inside the perimeter on the request path — controlling what reaches the model at all.
Cost as a security signal
Token and cost consumption is metered per interaction, and hard budgets are enforced while a response is still streaming — containing misuse, runaway loops and denial-of-wallet. The limit enforces; it does not merely alert.
Oversight that keeps pace
Policy violations surface in a live audit feed, and controls are adjusted centrally and take effect across the estate in seconds, without redeploying agents or changing connected systems.
Sovereign by placement
The control plane operates on-premises, in sovereign cloud or fully air-gapped, and can run with no external dependency at runtime and no telemetry leaving the environment.
Uniform across the estate
One policy definition and one evidence trail apply across participating, mediated harnesses and models, so adding a harness does not add a governance regime.
Answering the governance questions
The guidance closes with questions directors and executives should ask: what data, systems and tools can the agent access? What actions can it perform? Can all significant decisions and actions be monitored and audited? If the harness were compromised, misconfigured or manipulated, what would prevent or limit the outcome?
With controls implemented per harness, the answers are a collection of vendor attestations — one for each product, each on its own terms, none verifiable by the others.
With an independent control plane in the data path, the answers are evidence. What an agent could access is what the enforcement point permitted. What it did is what the record shows. And certain downstream effects of a compromised harness can be limited by a policy boundary the harness cannot reconfigure, where its traffic is mediated and alternative paths are controlled.
Defence in depth, not replacement
None of this replaces the controls inside a well-designed harness. Sandboxed execution, tool permissions, approval gates and context management belong in the harness, and ASD’s guidance describes them well. Which tools an agent may invoke, and what it may do inside connected systems, is decided there — the control plane governs what crosses the wire to the model, and evidences it.
The principle is the same one that runs through this research path: a component can secure its own service, but it should not be the enterprise’s only independent control over what that component may attempt. The harness governs the agent. The enterprise still needs to govern the interaction — independently, uniformly and with evidence it owns.
That is defence in depth, which is precisely what the guidance recommends.
Conclusion
ASD’s publication marks a shift: the layer above the model is now named, described and assigned responsibility by a national cyber-security authority. The conversation about agentic AI security has moved to where it belongs.
The enterprise conclusion follows directly. Models will keep changing. Harnesses will keep multiplying. The controls that ASD calls for — least privilege, complete records, independent monitoring, bounded cost, human accountability — need a home that does not move when a model is swapped or another agent product is adopted.
The harness is the layer above the model. The control plane is the layer above every harness.
And, as this path began: the control plane must be yours.
Independent, customer-controlled safeguards make valuable AI adoption easier to approve, operate and expand.
References
Australian Signals Directorate (2026). Agentic AI Harnesses: The layer above the model ↗. Published under Creative Commons Attribution 4.0.
Australian Signals Directorate (2026). Protect your organisations’ AI services ↗. Alert, 29 September 2026. Published under Creative Commons Attribution 4.0.
Australian Signals Directorate, with CISA, NSA, the Canadian Centre for Cyber Security, NZ NCSC and UK NCSC (2026). Careful adoption of agentic AI services ↗. Joint guidance, 1 May 2026.
UK National Cyber Security Centre (2026). Thinking carefully before adopting agentic AI ↗. Blog, 15 May 2026.
UK National Cyber Security Centre (2026). NCSC statement in response to recent incidents resulting from frontier AI evaluations ↗. Statement, 4 August 2026.
CERT-In (2026). Advisory CIAD-2026-0049: AI agents as attack vectors ↗. 15 October 2026.
Monetary Authority of Singapore (2026). Guidelines on Artificial Intelligence Risk Management for Financial Institutions ↗. 7 October 2026, effective 7 October 2027.
ISO/IEC (2026). ISO/IEC 27090:2026 ↗. Published 14 October 2026.
NIST (2026). Back to the Future: Why Agentic AI Needs a Strong Identity Foundation ↗. Blog, 27 August 2026.
TomorrowX (2026). The Control Plane Must Be Yours.
TomorrowX (2026). Data Mediation™ and Programmable Data Agents.