You Don’t Have to Be in Europe: The EU AI Act Compliance Clock Is Already Running
Dr. Anjan Krishnamurthy••
4 min read
Share:
Share:
Most security leaders read the EU AI Act as somebody else’s regulation — a European law for European companies. That reading is the single most expensive misconception in AI governance today. The Act regulates by risk, not by industry, and it follows the output of an AI system wherever that output lands. If the result is used inside the Union, the law can reach you — from Bengaluru, New York or anywhere else.
Here is the reframe: the deadline moved, but the burden of proof did not. The window to build defensible, attestable evidence is open now, and it is shorter than it looks.
What Actually Changed
The Omnibus is settled law, not a proposal. The AI portion of the Digital Omnibus was adopted as Regulation (EU) 2026/1744, in force from 27 July 2026 — it defers timing, not obligations.
The main high-risk deadline moved to 2 December 2027. Annex III high-risk systems shift from 2 August 2026 to 2 December 2027; AI embedded in regulated products (Annex I) moves to 2 August 2028.
Article 15 hardened into a security mandate. High-risk AI must be accurate, robust and resilient — the Act names the attack classes explicitly: data poisoning, model poisoning, adversarial examples and confidentiality attacks. Adversarial testing becomes a legal requirement, not a best practice.
Extraterritorial scope is written into the text. Under Article 2, providers and deployers in any country are in scope where the output of their AI system is used in the Union. Recital 22 makes the intent explicit — you cannot dodge the Act by hosting the system offshore.
What Did Not Move
The obligations themselves. The prohibited-use ban has applied since February 2025. Obligations for general-purpose AI models have applied since August 2025. Transparency rules and the enforcement machinery — the EU AI Office plus national market-surveillance authorities — are live now. And the penalty ceilings are untouched: up to €35 million or 7% of global turnover for banned uses; €15 million or 3% for high-risk, transparency or deployer breaches; €7.5 million or 1% for supplying incorrect or misleading information — whichever figure is higher.
The Contrarian Read
Fifteen months is not a reprieve — it is the conformity-assessment bottleneck written into the calendar. Each high-risk system needs a documented risk-management process, data-governance controls, technical documentation, automatic logging, designed-in human oversight and a conformity assessment before it can carry a CE marking and be registered in the EU database. None of that can be assembled in the final quarter before a deadline; it is months of engineering and governance evidence per system.
And most organisations are behind. A Deloitte survey of 500 German AI managers in September 2024 found 48.6% had not seriously engaged with implementing the EU AI Act, and only about a quarter had actively started preparing. No landmark AI Act fine has landed yet — the high-risk regime does not bite until December 2027 — but that is exactly the point. Enforcement is a window that is opening, and the organisations that wait for the first penalty to make headlines will be the ones scrambling to catalogue systems they should have mapped years earlier.
What Leaders Should Do Now
Inventory and classify every AI system as tier-0 (PrismDiscover). You cannot comply with obligations for systems you have not found. Build an authoritative inventory across models, agents, MCP servers, API endpoints and ML systems — especially the high-risk tools buried in HR, credit and biometric workflows — then classify each by risk tier and confirm your role (provider, deployer, importer).
Map controls to your AI GRC obligations (PrismGovern). Map your programme to ISO/IEC 42001 and the NIST AI Risk Management Framework for a defensible backbone that aligns with what the Act expects and produces the risk files, data-governance controls, documentation and human-oversight records you will need as evidence.
Red-team high-risk AI to Article 15 (PrismStrike). Adversarially test high-risk models against data poisoning, model poisoning and evasion, and keep the results as robustness evidence — turning ‘we are resilient’ from an assertion into tested, documented proof.
Watch the action stream at runtime (PrismObserve). Compliance is a live state, not a one-time audit. Hold drift, incident and rollback evidence continuously so your robustness posture stays attestable between assessments.
Future Outlook
2 December 2027: the high-risk regime begins to bite. Expect early enforcement to target the organisations that cannot produce a system inventory or a conformity assessment on request.
Through 2026–27: a second, broader Omnibus track covering data and cybersecurity is still being negotiated — more change is coming, not less, so build governance that flexes rather than a point-in-time checklist.
Beyond the EU: expect the Act’s risk-tiering and Article 15-style robustness duties to become a de facto template that other jurisdictions echo, making the evidence you build now reusable well outside Europe.
How SISA PRISM Closes the Gap
SISA PRISM turns the Act’s demands into an operating system for evidence. Inventory and AI asset discovery (PrismDiscover) finds and tier-0 catalogues every model, agent and MCP server across five domains. Governance and control mapping (PrismGovern) maps your controls to ISO/IEC 42001, the NIST AI RMF and 40+ AI GRC frameworks. AI red-teaming (PrismStrike) adversarially tests high-risk systems for exactly the attack classes Article 15 names. And runtime monitoring and observability (PrismObserve) keeps drift, incident and rollback evidence live — so ‘we are resilient’ becomes tested, documented, defensible proof.
Associate Director at SISA PRISM with over 18 years of experience in cybersecurity, research, and academia. He holds a Ph.D. in Computer Science and Engineering, specialising in network security, cryptology, and cyber forensics. With five patents and 75+ publications, he drives innovation by transforming advanced research into real-world cybersecurity solutions.
