Skip to main content

EU AI Act Compliance (Articles 12 & 14)

Overview: High-Risk Systems & Oversight​

The EU AI Act classifies certain Generative AI deployments as "high-risk," imposing stringent requirements around traceability, continuous monitoring, and human oversight. A core challenge for enterprise architects is achieving this traceability without violating concurrent data minimization mandates (The Article 12 Paradox).

The LLM-Shield-Proxy systematically addresses Articles 12 (Record-keeping) and 14 (Human oversight) through cryptographic attestation and hard systems-level containment.

Satisfying Article 12: Record-Keeping and Traceability​

Article 12 mandates that high-risk AI systems automatically record events ('logs') over their lifetime to ensure traceability of the system's functioning.

WORM Audit Logging & Merkle Chaining​

To provide absolute traceability without persisting raw user prompts to disk, the proxy implements WORM (Write Once, Read Many) Audit Logging.

  • SHA-256 Sequential Merkle Hash Chaining: Every redaction, tool-call interception, and configuration change generates a cryptographic event. These events are linked using sequential SHA-256 hashes, creating a Merkle chain. Any retroactive tampering with the logs will immediately invalidate the chain.
  • Proof of Non-Egress Receipt: The proxy computes a rolling SHA-256 digest over the entire SSE stream. It emits an HMAC-signed attestation proof guaranteeing exactly what data was (and wasn't) sent to the external LLM provider, providing mathematical proof to EU auditors.

NIST OSCAL Decision Traces​

The Universal Decision Trace Exporter formats these cryptographic events into automated NIST OSCAL (SP 800-53 Rev. 5) assessment results and OpenTelemetry gen_ai.* spans. This allows seamless ingestion into GRC systems (Vanta, Drata) for continuous, provable record-keeping.

Satisfying Article 14: Human Oversight and Agent Containment​

Article 14 dictates that high-risk systems must be designed to allow effective human oversight to prevent or minimize risks to health, safety, or fundamental rights.

Streaming Tool-Call RBAC​

As AI agents become autonomous, the risk of unauthorized lateral movement (e.g., executing malicious SQL or shell commands) increases exponentially.

  • Mid-Stream Interception: The proxy features pluggable streaming tool-call RBAC that intercepts JSON-RPC 2.0 / MCP (Model Context Protocol) function calls mid-stream.
  • Policy Evaluation: Tool calls (like exec_sql) are synchronously evaluated against OPA (Open Policy Agent) and HashiCorp Vault resolvers utilizing atomic dictionaries and thundering-herd locks.

Composite Agent Loop Circuit Breakers​

To prevent runaway autonomous loops—a critical risk in agentic architectures—the proxy implements Composite Agent Loop Circuit Breakers. If an agent begins rapidly iterating or executing repetitive, unverified tool calls without human-in-the-loop validation, the circuit breaker halts the execution.

(Reference the Architecture & Cryptographic Data Flow for deeper implementation details on the proxy's streaming capabilities).