Glossary

European regulation (2024) that classifies AI systems by risk level (unacceptable, high, limited, minimal) and imposes proportionate obligations: documentation, human oversight, robustness and transparency. High-risk systems (biometrics, credit, recruitment…) are subject to strict requirements before market placement.

France's national authority for information systems security, reporting to the Prime Minister. Publishes technical frameworks (PVID, PRIS, SecNumCloud, AI security, post-quantum migration), qualifies and certifies security products, supports essential entities under NIS2, and coordinates responses to major incidents. Its frameworks are de facto standards in public procurement and sector-specific requirements (finance, healthcare, defence).

CVE (Common Vulnerabilities and Exposures): unique identifier assigned to each known software vulnerability, the universal reference for patch management. CVSS (Common Vulnerability Scoring System): score from 0 to 10 assessing the criticality of a CVE based on exploitability and impact on confidentiality, integrity and availability. A CVSS score of 9 or above designates a critical vulnerability. CVE management is an explicit obligation under MCC and NIS2.

Digital Operational Resilience Act, European regulation (applicable since January 2025) imposing digital operational resilience requirements on financial entities (banks, insurers, PSPs): ICT risk management, resilience testing, major incident reporting and oversight of critical third-party providers.

Second European Payment Services Directive (2015, revision to PSD3/PSR in progress). Mandates strong customer authentication (SCA) for online payments, opens bank account data to authorised third parties (open banking) via standardised APIs, and strengthens PSP liability for fraud. The regulatory foundation for e-KYC and identity verification flows in the financial sector.

Revision of the European regulation on electronic identification and trust services. Requires each member state to deploy an interoperable digital identity wallet (EUDIW). Defines new assurance levels for qualified signatures, seals and timestamping. Relying parties (online services) must accept the wallet by 2026.

European Union Digital Identity Wallet, sovereign digital wallet deployed by each member state under eIDAS 2.0. Allows citizens to store verifiable credentials (PID: identity data, QEAA: diplomas, licences…) and present them selectively to online services via OpenID4VC/VP protocols.

Passwordless strong authentication standard driven by the FIDO Alliance. Based on asymmetric cryptography: a key pair is generated on the user's device, the private key never leaves the device. WebAuthn (W3C) is the associated browser API. Passkeys are the consumer implementation (synced via iCloud/Google).

Technique for adapting a pre-trained language model (LLM) on a specific corpus (instructions, business examples) to specialise its behaviour without retraining from scratch. To be distinguished from RAG: fine-tuning permanently modifies model weights (costly, risk of catastrophic forgetting of prior knowledge); RAG injects context at inference time (more agile, verifiable sources). Both approaches are complementary for business use cases.

Filters and constraints applied around a language model (LLM) to bound its behaviour in production: blocking out-of-scope topics, preventing confidential data leakage, detecting content drift, validating outputs before delivery. An essential component of an AI Act-compliant architecture.

Phenomenon in which a language model (LLM) generates factually incorrect information presented with confidence. Results from the probabilistic nature of models: they predict likely tokens, not verified facts. RAG architectures reduce hallucinations by grounding generation on controlled document sources. Detecting and limiting hallucinations is an explicit requirement of the AI Act for high-risk systems.

Certification issued by ANS (Agence du Numérique en Santé) for any provider hosting personal health data in France. Mandatory since 2018 for digital health applications (patient records, telemedicine, medical management software). Imposes security requirements close to SecNumCloud: France-based hosting, strict access controls, regular audits. Distinct from ISO 27001 certification, which it encompasses.

AI paradigm in which one or more language models (LLMs) autonomously orchestrate sequences of actions (tool calls, web browsing, code execution) to achieve a complex goal without human intervention at each step. Raises critical oversight challenges: cascading error amplification, decision traceability, expanded attack surface (prompt injection). Explicitly addressed by the AI Act for high-risk autonomous systems.

Know Your Customer, regulatory obligation (AML/CFT, financial regulators) to verify a customer's identity before entering into a business relationship. e-KYC (electronic KYC) performs this verification remotely: document control, liveness detection and biometric comparison. The French PVID framework is the regulated equivalent at eIDAS Substantial assurance level.

Anti-Money Laundering / Countering the Financing of Terrorism, regulatory framework (transposition of EU AML directives) imposing customer due diligence (KYC), suspicious transaction reporting and asset freezing obligations on obliged entities (banks, insurers, notaries, real estate agents…).

Set of practices for operating language models (LLMs) in production: model and prompt versioning, response quality monitoring, drift detection, update management without regression, observability and decision traceability. An extension of MLOps practices specific to generative models.

Critical system managed services contract structured around four axes: availability (defined and measured SLAs), security (patches, CVE management, progressive hardening), compliance (GDPR, AI Act, NIS2 on an ongoing basis) and evolvability (corrective and evolutionary maintenance). Distinct from traditional managed services in that it includes ongoing compliance responsibility.

Access control method requiring at least two factors of different types: something you know (password), something you have (OTP, hardware key) and something you are (biometrics). SMS-OTP is a weak factor, vulnerable to SIM swapping; FIDO2/Passkeys represents the MFA level most resistant to phishing attacks. Mandated by PSD2 (strong customer authentication, SCA) for payments, and recommended by ANSSI for access to critical systems.

European directive on the security of network and information systems (2022). Significantly broadens the scope of covered entities (essential and important entities) and strengthens obligations: cybersecurity governance, risk management, 24-hour incident reporting and supply chain security.

Open standards for authorization delegation (OAuth 2.0) and federated authentication (OpenID Connect, an identity layer on top of OAuth 2.0). Foundation of most enterprise SSO flows and third-party integrations. OpenID4VC (the variant used by EUDIW) extends the protocol to verifiable credential presentation. Distinct from FIDO2: handles identity federation, not first-factor authentication.

Set of components (certificate authorities, registration authorities, directories, CRL/OCSP) managing the lifecycle of X.509 digital certificates. The foundation of TLS/mTLS, code signing and qualified electronic signatures. An organisation's PKI defines the level of trust granted to its digital identities.

Attack targeting an LLM-based system by injecting, within the input data, instructions that override or replace the original system prompt. Example: a document submitted to an analysis agent contains a hidden instruction redirecting the agent to an unauthorised action. Distinct from data poisoning, which targets model training. Listed as a priority attack vector in the OWASP Top 10 for LLMs.

Attack targeting the training or fine-tuning phase of a model: malicious data is injected into the corpus to create a persistent undesired behaviour (backdoor, deliberate bias, information leak on trigger). Distinct from prompt injection, which is an inference-time attack. Particularly concerning for models trained on open sources or outsourced pipelines. A threat explicitly addressed by the AI Act in training data governance requirements.

Remote Identity Verification Provider, ANSSI (France) framework defining technical and organisational requirements for remote identity verification at eIDAS Substantial assurance level. Used for online bank account opening, gaming platforms and dematerialised notarial operations.

AI architecture that connects an LLM to an external document base via a vector search engine. The model retrieves relevant passages before generating its response, reducing hallucinations and enabling use of recent or proprietary data. A "sovereign" RAG architecture hosts the entire chain (model, vectors, data) on a controlled infrastructure.

General Data Protection Regulation, European regulation (2018) governing the collection and processing of personal data. Enshrines data subject rights (access, rectification, erasure, portability) and imposes on controllers documentation obligations (records of processing, DPIA), data minimisation, and breach notification within 72 hours. AI systems processing personal data are subject to the dual constraints of GDPR and the AI Act.

ANSSI qualification framework for cloud service providers. Requires immunity from extra-European law (Patriot Act, Cloud Act), localisation of data and operations in France or Europe, and high security requirements (encryption, access management, audit). The cloud baseline for critical systems, operators of essential services under NIS2, and the healthcare sector (HDS).

Structured inventory of all software components in a system: open source libraries, third-party dependencies, exact versions and licences. Enables real-time identification of components affected by a new CVE (e.g. Log4Shell) and quantification of the supply chain attack surface. Required by the NIS2 directive and the European Cyber Resilience Act (CRA) for software distributed on the European market.

Three levels defined by eIDAS. SES (Simple Electronic Signature) is any electronic data linked to a signatory (checkbox, signature image). AES (Advanced Electronic Signature) is uniquely linked to the signatory, capable of identifying them, and under their sole control. QES (Qualified Electronic Signature) is created using a qualified signature creation device (QSCD, e.g. HSM) with a qualified certificate issued by a QTSP; it carries the same legal value as a handwritten signature across the EU and is the only level automatically recognised without national equivalence.

Security Information and Event Management (SIEM): platform centralising the collection, correlation and real-time analysis of security logs and events. Security Operations Center (SOC): team and processes for security incident analysis and response, relying on the SIEM. Required for essential entities under NIS2, and a central component of MCC for intrusion detection and incident response within regulatory notification timeframes.

Progressive legacy system modernisation technique: the new system is developed in parallel with the old one, features are progressively migrated and traffic is redirected incrementally. The old system is retired once fully replaced. Avoids the "big bang" and guarantees service continuity during the transition.

W3C standard defining an interoperable format for digital credentials (degrees, licences, identity attestations) issued, held and selectively presented by their holder. Technical pillar of EUDIW: PIDs (identity data) and QEAAs (qualified attestations) are VCs conforming to the SD-JWT VC or MDOC profile. Digital counterpart of a paper document, cryptographically verifiable without querying the issuer.

Security architecture model based on the principle "never trust, always verify": no resource is considered safe by default, whether internal or external to the network perimeter. Every access is continuously authenticated, authorised, and logged. Relies on strong identity (FIDO2, mTLS), micro-segmentation, and continuous traffic inspection. Approach recommended by ANSSI and aligned with NIS2 and DORA requirements.