Cybersecurity Center

Cybersecurity and the Architecture of the Future

Artificial intelligence is transforming the IT landscape at an unprecedented pace. While many organizations are focused on productivity gains and eliminating manual work, a more fundamental transformation is taking place in the background: an asymmetric race against time in cybersecurity.

The urgency is clear: organizations that fail to adapt their security and architecture to this pace will be driven out of the market by cyber incidents or the inability to obtain insurance. To understand and control this transformation, we need to examine the three successive waves of the AI security challenge—and the architecture required to stay resilient.

The Three Waves of the AI Security Transformation

3 waves of cybercrime

Wave 1: Technical Vulnerabilities in Code (2025–2027)

  • The problem: The current explosion in data breaches and intrusions is the result of a historic catch-up effort. AI tools are scrutinizing decades of legacy code written by humans. Classic vulnerabilities—such as SQL injection, memory leaks, and outdated libraries—are being identified by attackers at scale.

  • The dynamic: This phase centers on “Shift Left” and automated code hygiene. Companies that clean up their pipelines and equip them with automated testing and patching can quickly make their traditional codebases more secure than ever.

  • The solution: Ongoing Automated Patching and Code Cleanup CI/CD pipelines are equipped with AI gatekeepers that continuously scrutinize codebases, automatically generate patches, and deploy them once integration tests have passed.

Wave 2: Abuse of Business Logic (2026–2028)

  • The problem: Once the technical front door is secured, the battle shifts to the functional logic. An API may be 100% secure from a technical perspective—free of SQL injection and buffer overflows—but if a user or script can manipulate the system through a specific sequence of actions, such as race conditions involving discounts, BOLA attacks, or skipping process steps, this can cause immediate financial damage.

  • The dynamic: Attackers no longer break the system; they make it operate in a way that costs the organization money. This results in continuous, invisible erosion of margins.

  • The solution: Formal Verification & Strict Finite-State Machines (FSMs).

    • Finite State Machines (FSMs): Business processes are not written as separate API endpoints, but mathematically modeled as a finite state machine. The system can only move from state A (CART_CREATED) to state B (PAYMENT_PENDING) and C (ORDER_COMPLETED). Manipulating or skipping steps is made technically impossible at the server level.

    • Idempotency and Atomic Transactions: Actions are made idempotent and processed using strict database isolation levels. Race conditions (submitting multiple requests simultaneously to stack discounts) are automatically blocked or placed in a queue at the database level.

    • AI-Assisted Formal Verification AI agents are deployed in the build pipeline to use mathematical verification to prove that the software can only assume the intended logical states, even before the code goes into production.

Wave 3: Manipulation of AI Agents and Decision-Making (2027–2030+)

  • The problem: In an fully agentic world, autonomous AI agents work together and make decisions based on data from the outside world (emails, invoices, documents). At this stage, attackers target Indirect Prompt Injection, Data Poisoning and taking over agent goals.

  • The dynamic: The source code and logic are correct, but the AI agent is misled by the context it reads. The question shifts from “Is the code secure?” to “Is the agent’s decision-making reliable?”.

  • The solution: The 3-Layer A2A Communication Model (Agent-to-Agent).

The Architectural Solution for Wave 3: The Three-Layer A2A Model

To allow AI agents to negotiate flexibly in complex situations and edge cases without exposing the organization to the risks of Wave 3, a layered communication architecture is essential. The principle is simple: you negotiate in natural language, but execute through a formal contract.

A2A architecture

The Three Layers of the A2A Architecture:

Layer 1: The Free Negotiation Layer (Expression & Creativity)

Here, agents communicate with one another in a rich context to explore complex issues, prices, and conditions.

  • The risk: This is the layer most vulnerable to influence, prompt injection, and manipulation.

  • The safeguards: At this layer, agents may no direct action whatsoever execute in the production infrastructure. It is an isolated sandbox. In addition, we monitor this layer with Cognitive Rate Limiting (preventing an agent from probing parameters without limit) and Cross-Model Validation (a second, differently trained AI model assesses the negotiation outcome for manipulation).

Layer 2: The Formalization and Intent Layer (Legal Safeguarding)

As soon as agents in Layer 1 reach agreement on a solution or transaction, the outcome is stripped of all free-form language and narrative.

  • The safeguards: The result is converted into a formal, structured contract (using a Domain-Specific Language or predefined JSON schema). This explicitly defines the obligations, limits, and conditions. This layer acts as the digital “notary” that verifies whether the proposal complies with the predefined business frameworks.

Layer 3: The Deterministic Execution Layer (The Unshakable Foundation)

This is the level at which the actual transaction or system change is implemented. No AI agents communicate here; instead, rigid mathematical systems communicate through the monitored logic from Wave 2.

  • The safeguards:

    • Cryptographic Verification: Messages between systems are secured using Mutual TLS (mTLS), equipped with unique nonces (to prevent replay attacks), and signed with cryptographic keys.

    • Deterministic Gateways: API gateways independently verify whether a request falls within the strict limits, such as a maximum transaction amount.

The New Role of Human Experts

This layered approach shows that the role of human specialists does not disappear, but shifts toward defining the system’s framework:

  • The Procurement Officer designs the strategies, parameters, and boundaries within which the agent in Layer 1 may negotiate.

  • The Legal Counsel designs the templates, ontologies, and formal logical frameworks for Layer 2.

  • The Security Architect builds the deterministic, monitored execution environment in Layer 3 and safeguards the FSM logic from Golf 2.

By strictly separating the “thinking and negotiation phase” from the “execution phase,” we retain the flexibility of AI agents while ensuring that control over business operations remains firmly in the hands of the organization.

Need help transforming your organization toward AI-driven cybersecurity and automation? Contact us today contact us

Gerard

Gerard works as an AI consultant and manager. With extensive experience at large organizations, he can quickly get to the root of a problem and work toward a solution. Combined with his background in economics, this enables him to make sound business decisions.