External governance layers are now required to meet new AI deployment liabilities
As new laws in Europe and California place strict legal liabilities on companies deploying artificial intelligence, industry reliance on internal model alignment is proving insufficient for regulatory compliance and audit trails.
Companies deploying artificial intelligence are discovering that internal model alignment cannot satisfy new legal and audit requirements. Regulators and auditors are asking for deployment records, not just assurances about how a model behaves.
This distinction creates a major compliance gap for businesses across Europe and beyond. The European Union’s AI Act places specific duties on deployers alongside obligations for labs, while data-protection laws hold organizations responsible for personal data usage.
Security engineers identified this exact problem fifty years ago. A 1972 US Air Force study established that a trusted reference monitor must be tamperproof, invoked on every access, and small enough to verify completely.
Modern frontier models fail all three conditions. They are vulnerable to prompt injection because instructions and data share the same text channel, they act as confused deputies when given tools, and their billions of opaque parameters cannot be audited like a small enforcement kernel.
The risk expands rapidly as models gain agency and execute actions through external tools. OWASP’s 2026 tracking reveals that prompt injection maps to six of the ten risks in its agentic top-ten list, showing the ungoverned surface grows alongside adoption.
Public benchmarks like StrongREJECT, HarmBench, and AgentHarm evaluate models on generic prompts rather than specific corporate policies or user data. A high safety score reflects a property of the model itself, not a usable audit trail for a specific deployment.
The solution requires building a thin control layer around the model rather than trying to align the weights perfectly. This external policy engine treats the model as an untrusted component and mediates every proposed action before it reaches a real resource.
When a model proposes an action, the external engine checks it against actual corporate policy and the specific capabilities granted for that task. The system then logs exactly who acted, which model was used, what data was involved, and what was decided.
This architecture restores the tamperproof, always-invoked, and verifiable properties that models lack. Research defenses like CaMeL, developed by Google DeepMind and ETH Zurich, already achieve provable security by containing the model within this exact type of external boundary.
Alignment remains necessary for shaping model disposition, but it answers the wrong question for regulators. Closing the gap between model safety and verifiable deployment governance is the actual requirement for businesses navigating the new legal landscape.
