EU AI Act Art. 15: accuracy, robustness and cybersecurity

Accuracy, robustness and cybersecurity

By Jorge García Aibar Reviewed 2026-09-24

From clause to evidence Build the artefact, then file every output it produces as a record that names AIGE-OBL-EUAIA-ART15. Drawn from chapter 08.

Text alternative

Clause: EU AI Act, Art. 15.
Duty holder: Provider.
Applies from: 2027-12-02, Deferred.
Artefact: Eval gate.
Layers: Layer 03, Layer 04.
Evidence record: Eval result, +4 more.
Record schemas: Eval result , Control observation , Test plan , Test report , Design record .
The same topic in 21 other frameworks; the crosswalk section below links each clause.

When this obligation applies The first date in the row's status, then each later step: filled once its date has passed, outlined while it is still ahead of the review date.

Data table: the dates

Dates of AIGE-OBL-EUAIA-ART15
Date	Step	Detail
2027-12-02	Deferred	Annex III
2028-08-02	High-risk (Annex I)	Applies to Annex I embedded (product safety-component) systems
2030-08-02	Later step	Deadline for legacy high-risk systems intended for use by public authorities (Art. 111(2))

Id

AIGE-OBL-EUAIA-ART15

Instrument

EU AI Act (post-Omnibus) law

Compared side by side

ISO 42001 vs EU AI Act · NIST AI RMF vs EU AI Act

Clause

Art. 15

Duty holder

Provider

Authority

National MSA

Applies from

2027-12-02 Deferred · Annex III

Later dates

2028-08-02 Applies to Annex I embedded (product safety-component) systems
2030-08-02 Deadline for legacy high-risk systems intended for use by public authorities (Art. 111(2))

System class

High-risk (Annex III) · High-risk (Annex I)

The artefact that evidences it

Eval gate; adversarial red-team suite; robustness and security controls; regression evals.

What this obligation connects to Solid edges are direct relations, dashed ones looser (a case that cites the same article, a framework whose clauses only touch the topic); patterns and controls take the colour of their stack layer. Select a node to open it; the lists below give every relation in full.

Data table: every relation

Relations of AIGE-OBL-EUAIA-ART15
Family	Item	Relation
Patterns	AI Threat Model	Direct
Patterns	Adversarial Red-Team Suite	Direct
Patterns	Agent Identity & Scoped Credentials	Direct
Patterns	Claims Substantiation Gate	Direct
Patterns	Drift & Fairness Monitor	Direct
Patterns	Eval Gate in CI	Direct
Patterns	Fairness Eval Suite	Direct
Patterns	Kill Switch / Circuit Breaker	Direct
Patterns	Model Artefact Integrity	Direct
Patterns	Runtime Guardrail	Direct
Open controls	AIGE-CTL-AGENT-025 Accountability across hops	Direct
Open controls	AIGE-CTL-AGENT-015 Checkpoints on irreversible actions, failing closed	Direct
Open controls	AIGE-CTL-AGENT-016 Code runs only in a sandbox	Direct
Open controls	AIGE-CTL-ASSURE-006 Control Observations Filed Against Control Ids	Direct
Open controls	AIGE-CTL-EVAL-003 Credential Isolation	Direct
Open controls	AIGE-CTL-AGENT-022 Delegation, never impersonation	Direct
Open controls	AIGE-CTL-EVAL-009 Evaluation Validity Checks	Direct
Open controls	AIGE-CTL-AGENT-008 Execution budgets	Direct
Open controls	AIGE-CTL-EVAL-008 Harness and Configuration Attestation	Direct
Open controls	AIGE-CTL-AGENT-002 Its own identity	Direct
Open controls	AIGE-CTL-AGENT-019 Local MCP servers sandboxed	Direct
Open controls	AIGE-CTL-AGENT-020 MCP authorisation (spec 2026-07-28)	Direct
Open controls	AIGE-CTL-AGENT-018 MCP server admission gate	Direct
Open controls	AIGE-CTL-AGENT-023 Memory write gate and rollback	Direct
Open controls	AIGE-CTL-ASSURE-010 Model Artefacts Signed at Build and Verified Before Load	Direct
Open controls	AIGE-CTL-EVAL-002 Network Egress Control	Direct
Open controls	AIGE-CTL-AGENT-017 Output and egress filter	Direct
Open controls	AIGE-CTL-AGENT-010 Per-agent circuit breaker	Direct
Open controls	AIGE-CTL-AGENT-032 Plan reviewed before it runs	Direct
Open controls	AIGE-CTL-ASSURE-002 Release Blocked Below the Eval Threshold	Direct
Open controls	AIGE-CTL-AGENT-021 Replace long-lived secrets with short-lived credentials	Direct
Open controls	AIGE-CTL-AGENT-039 Restrictions tested by attempting what is denied	Direct
Open controls	AIGE-CTL-AGENT-007 Runtime guardrail on every tool call	Direct
Open controls	AIGE-CTL-ASSURE-011 Safe Model Formats and Digest-Pinned Third-Party Models	Direct
Open controls	AIGE-CTL-AGENT-026 Stopping third-party agents at your boundary	Direct
Open controls	AIGE-CTL-AGENT-037 The person types the secrets	Direct
Open controls	AIGE-CTL-AGENT-005 Tool allow-list, deny by default	Direct
Open controls	AIGE-CTL-AGENT-036 Typed messages and limited shared memory between agents	Direct
Same crosswalk topic	AI Verify 4.8.1, 4.1.1, 6.1.1, 5.2.1, 5.9.1	Core clause
Same crosswalk topic	CSA AICM TVM-13, AIS-09, AIS-10, MDS-06, MDS-07, AIS-05	Core clause
Same crosswalk topic	China Deep Synthesis DeepSyn Art. 10, DeepSyn Art. 15, DeepSyn Art. 20	Core clause
Same crosswalk topic	China GenAI Measures GenAI Art. 10, GenAI Art. 14, GenAI Art. 17	Core clause
Same crosswalk topic	CoE Convention CoE Art. 16(2)(g)	Core clause
Same crosswalk topic	EU AI Act Art. 5, Art. 5(1)(a)–(b), Art. 55, Art. 60, Art. 9, Art. 42(3)	Core clause
Same crosswalk topic	G7 Code G7 Action 1	Core clause
Same crosswalk topic	GAO AI Accountability 3.7, 3.2	Core clause
Same crosswalk topic	GB/T 45654 GB/T 45654 Security assessment	Core clause
Same crosswalk topic	GPAI Code Safety C5, Safety 3.2, Safety C6	Core clause
Same crosswalk topic	ISO 42001 A.9, A.6, 9.1	Core clause
Same crosswalk topic	NIST AI RMF MANAGE 2, MEASURE 2, MEASURE 2.7, MEASURE 2.1, MEASURE 1	Core clause
Same crosswalk topic	OECD AI Principles OECD 1.4	Core clause
Same crosswalk topic	OWASP LLM LLM01:2026, LLM10:2026, LLM06:2026	Core clause
Same crosswalk topic	Singapore Agentic Agentic 2.3.1, Agentic 2.3.2	Core clause
Same crosswalk topic	Singapore GenAI GenAI 5, GenAI 6	Core clause
Same crosswalk topic	TC260 Framework 3.0 TC260 App. 2 II.5, TC260 3.2.1, TC260 3, TC260 App. 2 II.6, TC260 5.3.14	Core clause
Same crosswalk topic	China Algo. Rec. AlgoRec Art. 8, AlgoRec Art. 9	Related clause
Same crosswalk topic	GDPR Art. 32	Related clause
Same crosswalk topic	Korea AI Act Art. 32(1)	Related clause
Same crosswalk topic	OWASP Agentic ASI05	Related clause

Patterns that build it

Eval Gate in CI (layer 3)
Adversarial Red-Team Suite (layer 3)
Runtime Guardrail (layer 4)
Kill Switch / Circuit Breaker (layer 4)
Agent Identity & Scoped Credentials (layer 4)
AI Threat Model (layer 1 and 3)
Fairness Eval Suite (layer 3)
Model Artefact Integrity (layer 2 and 4)
Claims Substantiation Gate (layer 5 and 3)
Drift & Fairness Monitor (layer 4 and 5)

The same topic in other frameworks

From the topic crosswalk: the clauses filed under the same topics as this one. Mappings are illustrative, not a claim of conformity.

Runtime guardrails

ISO 42001 A.9 Use of AI systems (core)
NIST AI RMF MANAGE 2 MANAGE 2: Strategies to maximize AI benefits and minimize negative impacts are planned, prepared, implemented, documented, and informed by relevant AI actors (core)
TC260 Framework 3.0 TC260 App. 2 II.5 Dynamic runtime management (core)
TC260 Framework 3.0 TC260 3.2.1 Technological countermeasures for agentic AI (core)
China GenAI Measures GenAI Art. 10 Guided, bounded use (core)
China GenAI Measures GenAI Art. 14 Stop unlawful generation (core)
China Deep Synthesis DeepSyn Art. 10 Input and output review (core)
CSA AICM TVM-13 Guardrails (core)
CSA AICM AIS-09 Input Validation (core)
CSA AICM AIS-10 Output Validation (core)
OWASP LLM LLM01:2026 Prompt Injection (core)
OWASP LLM LLM10:2026 Improper Output Handling (core)
Singapore Agentic Agentic 2.3.1 During design and development, use technical controls (core)
EU AI Act Art. 5 Prohibited AI practices
ISO 42001 A.6 AI system life cycle
China Algo. Rec. AlgoRec Art. 8 Periodic algorithm review
China Algo. Rec. AlgoRec Art. 9 Feature database for unlawful content
EU AI Act Art. 5(1)(a)–(b) Manipulative techniques; exploitation of vulnerabilities
GPAI Code Safety C5 Commitment 5: Safety mitigations
OWASP LLM LLM06:2026 Unbounded Consumption
AI Verify 4.8.1 Safety: assess harms and implement measures that prevent the model generating unlawful or harmful content (CSAM, NCII, CBRN, offensive cyber)

Robustness, security and evaluations

EU AI Act Art. 55 Obligations for providers of general-purpose AI models with systemic risk (core)
ISO 42001 A.6 AI system life cycle (core)
NIST AI RMF MEASURE 2 MEASURE 2: AI systems are evaluated for trustworthy characteristics (core)
TC260 Framework 3.0 TC260 3 Technological countermeasures (core)
TC260 Framework 3.0 TC260 App. 2 II.6 Sandbox validation and red teaming (core)
China Deep Synthesis DeepSyn Art. 15 Technology management and algorithm verification (core)
China Deep Synthesis DeepSyn Art. 20 Security assessment of new products (core)
GB/T 45654 GB/T 45654 Security assessment Security-assessment requirements for generative AI services (core) (clause not verified)
GPAI Code Safety 3.2 Measure 3.2: Model evaluations (core)
NIST AI RMF MEASURE 2.7 MEASURE 2.7: AI system security and resilience as identified in the MAP function are evaluated and documented (core)
CSA AICM MDS-06 Adversarial Attack Analysis (core)
CSA AICM MDS-07 Robustness against Adversarial Attack / Model Hardening (core)
Singapore GenAI GenAI 5 Testing and Assurance (core)
Singapore GenAI GenAI 6 Security (core)
Singapore Agentic Agentic 2.3.2 Before deploying, test agents (core)
CoE Convention CoE Art. 16(2)(g) Testing before first use and when significantly modified (core)
OECD AI Principles OECD 1.4 Robustness, security and safety (core)
G7 Code G7 Action 1 Identify, evaluate and mitigate risks across the lifecycle, including testing (core)
GAO AI Accountability 3.7 Assessment: assess performance against defined metrics to ensure the AI system functions as intended and is sufficiently robust (core)
AI Verify 4.1.1 Safety (technical test): run use-case benchmarks and red-teaming to evaluate for safety and possible harms (e.g. hallucination, toxicity) (core)
AI Verify 6.1.1 Robustness (technical test): compare accuracy with and without added noise to document and monitor performance (core)
AI Verify 5.2.1 Security: conduct security risk assessments that model threats and risks to the AI system (core)
EU AI Act Art. 60 Testing of high-risk AI systems in real-world conditions outside AI regulatory sandboxes
ISO 42001 9.1 Monitoring, measurement, analysis and evaluation
TC260 Framework 3.0 TC260 5.3.14 Resilience
China GenAI Measures GenAI Art. 17 Security assessment
EU AI Act Art. 9 Risk management system
EU AI Act Art. 42(3) Presumption of conformity for cybersecurity (Cyber Resilience Act)
GPAI Code Safety C6 Commitment 6: Security mitigations
GDPR Art. 32 Security of processing
NIST AI RMF MEASURE 2.1 MEASURE 2.1: Test sets, metrics, and details about the tools used during TEVV are documented
NIST AI RMF MEASURE 1 MEASURE 1: Appropriate methods and metrics are identified and applied
CSA AICM AIS-05 Application Security Testing
OWASP LLM LLM01:2026 Prompt Injection
OWASP Agentic ASI05 Unexpected Code Execution (RCE)
Korea AI Act Art. 32(1) Safety duties for AI above the compute threshold
GAO AI Accountability 3.2 Metrics: define performance metrics that are precise, consistent, and reproducible
AI Verify 5.9.1 Security: release AI systems only after appropriate and effective security checks and evaluation

Open controls that evidence it

Draft controls in the open control profiles that map to this row: each states a requirement and the evidence it must leave behind.

AIGE-CTL-EVAL-002 Network Egress Control (Evaluation environment profile)
AIGE-CTL-EVAL-003 Credential Isolation (Evaluation environment profile)
AIGE-CTL-EVAL-008 Harness and Configuration Attestation (Evaluation environment profile)
AIGE-CTL-EVAL-009 Evaluation Validity Checks (Evaluation environment profile)
AIGE-CTL-AGENT-002 Its own identity (Agent runtime profile)
AIGE-CTL-AGENT-005 Tool allow-list, deny by default (Agent runtime profile)
AIGE-CTL-AGENT-007 Runtime guardrail on every tool call (Agent runtime profile)
AIGE-CTL-AGENT-008 Execution budgets (Agent runtime profile)
AIGE-CTL-AGENT-010 Per-agent circuit breaker (Agent runtime profile)
AIGE-CTL-AGENT-015 Checkpoints on irreversible actions, failing closed (Agent runtime profile)
AIGE-CTL-AGENT-016 Code runs only in a sandbox (Agent runtime profile)
AIGE-CTL-AGENT-017 Output and egress filter (Agent runtime profile)
AIGE-CTL-AGENT-018 MCP server admission gate (Agent runtime profile)
AIGE-CTL-AGENT-019 Local MCP servers sandboxed (Agent runtime profile)
AIGE-CTL-AGENT-020 MCP authorisation (spec 2026-07-28) (Agent runtime profile)
AIGE-CTL-AGENT-021 Replace long-lived secrets with short-lived credentials (Agent runtime profile)
AIGE-CTL-AGENT-022 Delegation, never impersonation (Agent runtime profile)
AIGE-CTL-AGENT-023 Memory write gate and rollback (Agent runtime profile)
AIGE-CTL-AGENT-025 Accountability across hops (Agent runtime profile)
AIGE-CTL-AGENT-026 Stopping third-party agents at your boundary (Agent runtime profile)
AIGE-CTL-AGENT-032 Plan reviewed before it runs (Agent runtime profile)
AIGE-CTL-AGENT-036 Typed messages and limited shared memory between agents (Agent runtime profile)
AIGE-CTL-AGENT-037 The person types the secrets (Agent runtime profile)
AIGE-CTL-AGENT-039 Restrictions tested by attempting what is denied (Agent runtime profile)
AIGE-CTL-ASSURE-002 Release Blocked Below the Eval Threshold (Assurance and evidence profile)
AIGE-CTL-ASSURE-006 Control Observations Filed Against Control Ids (Assurance and evidence profile)
AIGE-CTL-ASSURE-010 Model Artefacts Signed at Build and Verified Before Load (Assurance and evidence profile)
AIGE-CTL-ASSURE-011 Safe Model Formats and Digest-Pinned Third-Party Models (Assurance and evidence profile)

Source

Chapter 08, section EU AI Act, post-Omnibus, checked against its sources on the review date above.

Machine-readable

This obligation: /api/v1/obligations/aige-obl-euaia-art15.json
The register: /api/v1/obligations.json · CSV
Schema and stability promise: open data and API

Cite this obligation

García Aibar, J. (2026). EU AI Act Art. 15: accuracy, robustness and cybersecurity (AIGE-OBL-EUAIA-ART15). In AI Governance Engineering: The Thesis & Body of Knowledge (v0.5.0). https://doi.org/10.5281/zenodo.22956197. https://aigovernanceengineer.com/obligations/aige-obl-euaia-art15. CC BY 4.0

BibTeX

@misc{aige2026obligation,
  author       = {Jorge García Aibar},
  title        = {{EU AI Act Art. 15: accuracy, robustness and cybersecurity (AIGE-OBL-EUAIA-ART15)}},
  howpublished = {In AI Governance Engineering: The Thesis \& Body of Knowledge},
  year         = {2026},
  version      = {0.5.0},
  doi          = {10.5281/zenodo.22956197},
  url          = {https://aigovernanceengineer.com/obligations/aige-obl-euaia-art15},
  note         = {Version 0.5.0}
}