Skip to content

EU AI Act Compliance & High-Risk Cybersecurity (Article 15)

Density:

๐Ÿ›๏ธ High-Risk AI System Qualification (Annex III & Article 6)

Section titled โ€œ๐Ÿ›๏ธ High-Risk AI System Qualification (Annex III & Article 6)โ€

An AI system falls under mandatory High-Risk compliance if:

  1. Article 6 Harmonization Legislation: It serves as a safety component or is itself a regulated product under EU directives (e.g., Medical Devices, Machinery, Automotive, Aviation).
  2. Annex III Standalone Deployments: It operates in sensitive critical domains:
    • Category 2: Critical infrastructure management and safety components (energy, digital telecom, water, transport).
    • Category 5: Access to essential private and public services (credit scoring, insurance risk tiering, priority emergency dispatch).
    • Category 1: Remote biometric identification and physical/logical access control systems.

๐Ÿ›ก๏ธ Interactive EU AI Act Article 15 Assessment Engine

Section titled โ€œ๐Ÿ›ก๏ธ Interactive EU AI Act Article 15 Assessment Engineโ€

Use this diagnostic workbench to evaluate your models, agentic workflows, and LLM integrations against the technical mandates of the EU AI Act:

Section 84 โ€” EU AI Act Compliance Engine

High-Risk AI System Cybersecurity & Robustness Self-Assessment

Evaluate your AI models, agents, and pipelines against mandatory requirements of Regulation (EU) 2024/1689 (specifically Article 15 Cybersecurity & Robustness).

1

Step 1: High-Risk Use Case Qualification (Annex III & Article 6)

Select the operational category and mission profile of your AI deployment to identify statutory conformity requirements.

2

Step 2: Technical Cybersecurity & Robustness Diagnostic

Rate your system against core Article 15 requirements, prompt injection defenses, data poisoning shields, and fail-safe mechanisms.

Article 15(2) Adversarial Prompt Injection & Input Shielding

Does your AI system implement semantic guardrails and AST sanitizers preventing prompt overrides and tool-hijacking?

Hermes Solution: Section 75 Semantic Intent Divergence Tensor & Section 64 AST sanitizer.
Article 15(2) Training Data & Knowledge Base Poisoning Resistance

Are training datasets, RAG vector stores, and fine-tuning corpuses cryptographically notarized to detect backdoor tampering?

Hermes Solution: Section 74 Proof-of-Telemetry & Section 82 RFC 6962 append-only Merkle ledger.
Article 15(3) Deterministic Fallback & Graceful Degradation

Can the system fail-safe to deterministic heuristic rules with automated rollback if neural model confidence drops or detects anomalies?

Hermes Solution: Section 75 3PC Transactional State Machine & Section 70 active rollback.
Article 12(1) Tamper-Evident Post-Market Audit Trails

Are all model predictions, prompt inputs, decision rationales, and telemetry events logged into an immutable, verifiable ledger?

Hermes Solution: Section 82 RFC 6962 Continuous Epistemic Auditing & Section 79 DEC-* ledgers.
Article 14(4) Human Intervention, Kill-Switch & Automation Bias Mitigations

Can human operators immediately inspect the uncertainty bounds of AI outputs and physically trigger an emergency kill switch?

Hermes Solution: Section 83 Epistemic Limits ("Show Your Work") & Section 70 Dual-Key Kill Switch.
Article 9(2) Iterative AI Risk Assessment & Threat Modeling

Does your organization model multi-tier adversarial threats (STRIDE/PASTA/DREAD) specifically targeting AI model components?

Hermes Solution: Section 51 STRIDE Threat Modeling & Section 53 DREAD/FAIR quantitative scoring.

๐Ÿ” Technical Mapping: Article 15 vs. Hermes Codex Architecture

Section titled โ€œ๐Ÿ” Technical Mapping: Article 15 vs. Hermes Codex Architectureโ€

Hermes Codex directly provides the technical capabilities required to demonstrate Article 15 compliance:

Article 15.1 โ€” Byzantine Robustness

Statutory Mandate: High-risk AI systems shall be resilient against attempts by unauthorized third parties to alter their use or performance.

Hermes Implementation: Byzantine Fault Tolerant (PBFT / Raft-BFT) swarm consensus (Sections 71-73) tolerating up to f = floor((N-1)/3) compromised nodes, verified inside hardware confidential computing enclaves (AMD SEV-SNP, Intel SGX).

Article 15.2 โ€” Adversarial Shields

Statutory Mandate: Technical solutions addressing AI-specific vulnerabilities such as data poisoning, prompt injection, model evasion, and extraction.

Hermes Implementation: Section 75 Semantic Intent Divergence Tensor (tau_cos < 0.22), AST token stripping, and Section 74 Proof-of-Telemetry cryptographic anchoring.

Article 15.3 โ€” Failsafe & Graceful Degradation

Statutory Mandate: Redundancy, backup plans, and graceful degradation ensuring the system does not cause catastrophic failures during anomalies.

Hermes Implementation: Section 75 non-blocking 3PC transactional state machine with automated canary health rollback and fallback to deterministic rule engines.

Article 12 โ€” Verifiable Audit Trails

Statutory Mandate: Automatic logging over the AI lifetime to ensure traceability and facilitate post-market monitoring.

Hermes Implementation: RFC 6962 append-only Merkle tree logging (Section 82) providing logarithmic inclusion proofs (O(log2 N)) preventing retroactive modification.


Hermes Codex provides verified defenses against the 6 primary AI adversarial threat vectors:

  1. VEC-AIACT-01 (Prompt Injection & Jailbreak): Guarded via dual-phase semantic distance validation and AST parameter filtering.
  2. VEC-AIACT-02 (Data & Embedding Poisoning): Guarded via SHA-256 Merkle proofs anchoring all ingested training and RAG documents.
  3. VEC-AIACT-03 (Model Evasion & Perturbations): Benchmarked continuously via automated Breach and Attack Simulation (BAS) fuzzing.
  4. VEC-AIACT-04 (Model Inversion & Weight Theft): Isolated in memory via hardware CPU enclaves (AMD SEV-SNP / Intel SGX).
  5. VEC-AIACT-05 (Unconstrained Agentic Escalation): Controlled via Two-Phase Commit (2PC) state machines requiring dual authorization for destructive actions.
  6. VEC-AIACT-06 (Epistemic Hallucination): Mitigated by explicit UNCOMPUTABLE state tagging under Section 83 methodology.