We are attempting to regulate a dynamic, emergent technology using static, paper-based workflows. For the CISO, the EU AI Act is a necessary foundation—but treating compliance as a complete defense against autonomous AI is a dangerous illusion.
Consider a hypothetical scenario that keeps security leaders awake.
In the spring of 2026, a regional energy grid in Northern Europe deploys an autonomous AI routing system to optimize power distribution. Three months later, the system reroutes energy in a way human operators did not anticipate, isolating three substations for forty-seven minutes.
When the regulators arrive, they ask a simple question:
Why did the AI make that decision?
The energy company’s governance team produces a 400-page Fundamental Rights Impact Assessment. It demonstrates that the system was assessed and classified appropriately at the time of deployment. But it cannot reconstruct the precise chain of computational and environmental events that led to the decision on that Tuesday afternoon.
The system has changed. Its environment has changed. Its interactions with other systems have changed.
The compliance document remains intact.
The operational reality does not.
As we convene for the Amsterdam Edition on November 12th, this scenario represents one of the defining challenges of 2026. The European Union has created one of the world’s most ambitious frameworks for governing artificial intelligence. But enterprises are beginning to confront a fundamental distinction:
Regulatory compliance is not the same thing as runtime control.
The Three Layers of AI Security
To understand the challenge, we must separate AI risk into three distinct layers.
Regulatory Governance: What the EU requires through frameworks such as the AI Act.
Technical Governance: How organizations architect, test, validate, monitor, and control AI systems before and after deployment.
Runtime Security: What happens when an autonomous system is operating in a live, complex, and potentially hostile environment.
The EU AI Act establishes an important governance baseline. For high-risk systems, it introduces requirements around risk management, data governance, technical documentation, record-keeping, human oversight, accuracy, robustness, and cybersecurity.
But regulation cannot by itself provide continuous assurance over a dynamic autonomous system.
The fundamental problem is not that the legislation is irrelevant. It is that legislation and runtime behavior operate on different timescales.
AI models can be updated. Data distributions can change. Dependencies can fail. Agents can interact with one another. New attack techniques can emerge after deployment.
A model integrated into a financial environment may present one risk profile in January and a materially different operational risk by August because of changes in data, integrations, permissions, or adversarial behavior.
Compliance is necessary. Compliance alone is not continuous security.
The Complexity of the Compliance Landscape
This governance challenge is being compounded by the evolving European regulatory landscape.
The ongoing political debate around proposed changes to AI legislation and implementation timelines has created additional uncertainty for security leaders. Policymakers are attempting to balance regulatory certainty with the need to avoid slowing innovation, while enterprises are simultaneously under pressure to deploy AI at unprecedented speed.
For CISOs, the practical challenge is straightforward:
Technology is moving faster than the governance cycle.
Organizations cannot simply wait for every regulatory question to be resolved before implementing meaningful security controls.
The challenge becomes even more complicated when AI is embedded into critical infrastructure.
The AI Act’s interaction with existing sector-specific and product-safety frameworks creates a complex regulatory landscape for systems operating in environments such as energy, transport, healthcare, and industrial control.
The real danger is not necessarily a legal exemption.
It is the operational reality that when autonomous AI becomes embedded inside a complex physical system, the consequences of its decisions extend far beyond the AI component itself.
If multiple autonomous systems interact, a failure can propagate across dependencies faster than conventional governance processes can respond.
The question therefore becomes bigger than:
“Is this AI system compliant?”
It becomes:
“Can we observe, constrain, and recover from its behavior while it is operating?”
The Cryptographic Substrate
Even if regulatory frameworks were perfectly aligned, AI governance ultimately depends on another layer that receives far less attention: cryptographic infrastructure.
You cannot have AI sovereignty without cryptographic sovereignty.
This is not simply a geopolitical statement. It is a technical reality.
AI governance depends on trustworthy identity, authentication, integrity, provenance, access control, and auditability.
Every model signature, every authenticated human intervention, every security event, and every audit record relies on cryptographic mechanisms that establish trust between systems.
At the same time, CISOs are beginning to map their environments for the post-quantum transition.
The concern is not that today’s RSA and elliptic-curve cryptography has suddenly become obsolete. The concern is that sufficiently capable quantum computers could eventually undermine widely deployed public-key cryptography—and sensitive information encrypted today may already be attractive to adversaries pursuing “harvest now, decrypt later” strategies.
That creates a difficult question for AI governance:
What value does an immutable audit trail have if the cryptographic foundations protecting its authenticity and confidentiality eventually become vulnerable?
AI security and post-quantum security are therefore not isolated strategic problems.
They are increasingly connected.
The Epistemic Lattice: Moving Beyond the Audit
The current model—where a vendor documents a model, an enterprise completes its governance obligations, and the resulting documentation becomes evidence of compliance—must evolve.
To address the runtime security layer, we propose a structural concept we call the Epistemic Lattice.
The idea is simple:
Instead of treating governance as an external, post-hoc audit process, embed observation, authorization, and constraint directly into the infrastructure surrounding the AI system.
The objective is not to mathematically prove everything an AI “thought.” For complex neural networks, that remains beyond what current technology can reliably provide.
The objective is to control what the system is permitted to do.
An Epistemic Lattice would combine technologies and architectural principles such as:
- policy enforcement points;
- trusted execution environments;
- cryptographic identity and attestation;
- continuous runtime monitoring;
- tamper-evident logging;
- model and data provenance;
- explicit authorization boundaries;
- and post-quantum-ready cryptographic foundations.
The system would not simply ask an AI to explain itself after something goes wrong.
It would establish enforceable boundaries around the actions the AI can take, continuously observe its environment, authenticate critical interactions, and maintain a tamper-evident record of relevant inputs, decisions, authorizations, and outputs.
The fundamental shift is from auditing behavior after the fact to constraining behavior while the system is operating.
From reading the output to enforcing the boundaries of the computation.
The CISO’s Real Mandate
The problem isn’t that the EU AI Act is too weak.
The problem is that compliance cannot substitute for runtime control.
Boards must move beyond the assumption that satisfying regulatory requirements automatically means an enterprise is secure against autonomous systems.
Regulation, technical governance, and infrastructure controls must operate together.
The CISO’s responsibility is therefore evolving once again.
It is no longer enough to ask:
“Have we documented the risk?”
The more important questions are:
Can we observe the system?
Can we constrain it?
Can we prove who authorized its actions?
Can we detect when its operating environment changes?
And can we stop it when its behavior moves beyond acceptable boundaries?
At the Amsterdam Edition, we will move beyond the paper shields and examine the infrastructure beneath AI governance.
We will explore how organizations can build runtime security architectures, strengthen the cryptographic foundations beneath their AI systems, and implement dynamic risk controls that remain effective even as models, data, dependencies, and threats evolve.
Governance is not a document you file with a regulator.
It is a set of controls you enforce in your infrastructure.
Stop Auditing the Black Box. Start Constraining It.
Join senior CISOs, AI security leaders, technology executives, and cybersecurity decision-makers at the NEXT IT Security Amsterdam Edition on November 12th for a deep dive into AI governance, runtime security, cryptographic resilience, and the infrastructure required to secure autonomous systems.
