Cybersecurity and the architecture of the future

Artificial intelligence is rapidly transforming the IT landscape. While many organizations are focused on productivity gains and eliminating manual work, a more fundamental transformation is unfolding 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 forced out of the market by cyber incidents or the inability to obtain insurance. To understand and manage this transformation, we need to examine the three successive waves of the AI security challenge—and the architecture required to withstand them.

The Three Waves of the AI Security Transformation

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

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 Classic vulnerabilities—such as SQL injection, memory leaks, and outdated libraries—are being discovered by attackers at scale.

• The dynamic: This phase revolves around “Shift Left” and automated code hygiene. Companies that clean up their pipelines and equip them with automated tests and patches can make their traditional codebases safer than ever—quickly.

• The solution: Continuous 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 secure, the battle shifts to the functional logic. An API can be 100% secure from a technical perspective—without SQL injection or 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, direct financial damage results.

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

• The solution: Formal Verification & Strict State Machines (FSM).

  • Finite State Machines (FSMs): Business processes are not written as isolated 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.

  • Formal verification using AI: AI agents are deployed in the build pipeline to use mathematical verification to prove that the software can assume only the intended logical states, even before the code reaches production.

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

• The problem: In a 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 Architecture Solution for Wave 3: The Three-Layer A2A Model

To enable 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: negotiate in natural language, but execute through a formal contract.

The three layers of the A2A architecture:

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

At this layer, 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 safeguard: 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 indefinitely) and Cross-model validation (a second, differently trained AI model assesses the negotiation outcome for manipulation).

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

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

• The safeguard: The result is converted into a formal, structured contract (via a Domain-Specific Language or defined 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 carried out. No AI agents communicate here; instead, rigid mathematical systems communicate through the monitored logic from Wave 2.

• The safeguard:

  • 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 check, without AI intervention, 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 to defining the system’s parameters:

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

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

• The Security Architect builds the deterministic, controlled execution environment in Layer 3 and safeguards the FSM logic from Wave 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 with the transformation to AI-driven cybersecurity and automation? Contact us today contact touch