EU AI Act and LLM Proxies - Where VoidLLM Stands

The EU AI Act is rolling out in stages. Prohibited practices are in effect since February 2025, GPAI rules since August 2025. High-risk AI system obligations apply from August 2026, with full enforcement across all risk categories in August 2027.

What does this mean for the proxy layer sitting between your apps and the models? We spent time understanding this for VoidLLM. Here’s where we stand, as of Q2 2026.

Responsibility is layered, not binary

Under the AI Act, responsibilities are split across the value chain:

graph TD
  U[End User] -->|request| A[Your Application]
  A -->|API call| P[Proxy - VoidLLM]
  P -->|forward| M[Model Provider]
  M -->|response| P
  P -->|stream back| A

  P -.->|metadata only| L[(Usage Log)]

  style U fill:#1a1a24,stroke:#333,color:#e2e8f0
  style A fill:#1a1a24,stroke:#22c55e,color:#e2e8f0
  style P fill:#8b5cf6,stroke:#6366f1,color:#fff
  style M fill:#1a1a24,stroke:#333,color:#e2e8f0
  style L fill:#12121a,stroke:#8b5cf6,color:#e2e8f0

VoidLLM doesn’t train models or determine how outputs are used. Our responsibility is secure, well-documented infrastructure. If you expose AI capabilities to end users, the use case, risk classification, and transparency are yours.

That said, we’re not a completely neutral pipe. RBAC controls who accesses which models, rate limits constrain usage, routing determines which provider handles a request. These are infrastructure decisions, not AI decisions - but they shape the compliance landscape.

The legal classification of AI infrastructure providers is not yet fully settled. This is our position.

The prompt logging problem

Most LLM gateways log prompts and responses for debugging, analytics, or fine-tuning. Some do it by default.

Under GDPR, prompt content is likely personal data the moment a user types a name, email, or anything identifiable. If your proxy logs that:

• You need a legal basis to process it
• You need to include it in your records of processing activities
• You need to handle data subject access requests
• You need to delete it on request
• A data breach includes that content

This is active data minimization under GDPR Article 5(1)(c) and Privacy by Design under Article 25.

What the AI Act expects

Risk classification depends on the use case, not the proxy. Using GPT-4o for internal summarization is different from using it for credit scoring. The proxy doesn’t change the classification.

Transparency obligations. If end users interact with AI-generated content, they may need to be informed. This is an application-level concern.

Logging and monitoring. Article 12 requires “automatic recording of events” for traceability - primarily for high-risk systems. This targets metadata (who, when, which model, how many tokens), not conversation content.

In certain high-risk deployments, more detailed documentation may be needed for incident investigation or audit trails. But even then, that obligation sits with the deployer’s application architecture - the proxy doesn’t have the context to make that call.

What VoidLLM gives you

Without any enterprise license:

• RBAC - org/team/user/key hierarchy, model access allowlists, most-restrictive-wins
• Usage tracking - every request tracked with metadata
• Rate limits and token budgets - per org, team, and key
• Prometheus metrics - for monitoring and alerting
• Health probes - per-model health checking with circuit breakers

With enterprise license:

• Audit logs - every admin action logged, filterable, exportable
• OpenTelemetry tracing - request correlation across your infrastructure
• SSO/OIDC - centralized authentication

What VoidLLM doesn’t do

• We don’t validate AI outputs. Content filtering is the model provider’s or your responsibility.
• We don’t detect prohibited use cases. That’s governance, not infrastructure.
• We don’t inject AI disclosure labels. Your app handles user-facing transparency.
• We don’t assess risk levels. The use case determines the classification, not the proxy.

The compliance cost of content logging

When your proxy logs prompt content, the GDPR obligations stack up: processing records, retention policies, data subject requests, breach notifications, cross-border transfer rules, and a DPA covering content specifically.

With VoidLLM, none of that applies to the proxy layer. The compliance surface area is metadata only.

Some regulated industries may actively want content logging for their own reasons - that’s a legitimate choice. The point is it should be a conscious decision based on risk assessment, not a proxy default.

Disclaimer

We’re engineers, not lawyers. VoidLLM was designed to minimize the compliance surface area - not because regulation forced us, but because zero-knowledge is the right architecture for a proxy that handles other people’s data.

If you’re evaluating proxies for a regulated environment, the architecture matters more than the feature list.