Skip to content

AI Red-Team Plan Template (EU AI Act Article 15 & 55)

Template11 November 2026· 6 min read

A red-team plan template for EU AI Act Article 15 robustness and cybersecurity, and the Article 55 adversarial-testing duty for systemic-risk GPAI models.

Red-teaming is the practice of attacking your own system before someone else does. It is not a statutory term in the EU AI Act, Regulation (EU) 2024/1689 — but a structured adversarial-testing programme is how a provider generates evidence for two real duties. For any high-risk system, it feeds Article 15, which requires an appropriate level of robustness and cybersecurity, including resilience against attempts to alter use or performance by exploiting vulnerabilities. For providers of general-purpose AI models with systemic risk, Article 55 goes further: it makes adversarial testing a named obligation. This template covers both, because the discipline is the same even though the legal hook differs.

A red-team plan answers four questions in writing: what you are attacking (scope), how you will attack it (scenarios), how bad each finding is (severity), and what you will do about it (remediation). Without those four, "we red-teamed it" is a claim, not evidence.


It matters which duty your red-teaming serves, because the obligations and the actors differ:

High-risk AI systemSystemic-risk GPAI model
Legal hookArticle 15 robustness & cybersecurityArticle 55 (incl. adversarial testing)
WhoProvider of the high-risk systemProvider of a GPAI model meeting the Article 51 systemic-risk threshold
What red-teaming provesResilience to manipulation, errors, attacksModel evaluation, systemic-risk assessment and mitigation
Status (late 2026)High-risk requirements tied to the high-risk dateIn force — GPAI obligations applied from 2 Aug 2025

The timing contrast is important. The high-risk Article 15 duty is tied to the high-risk application date — originally 2 August 2026, moved to 2 December 2027 for Annex III systems by the Digital Omnibus, Regulation (EU) 2026/1744 (published in the Official Journal on 24 July 2026, in force since 27 July 2026). The Article 55 systemic-GPAI obligations, by contrast, are already in application (Chapter V applied from 2 August 2025 and was not deferred). If you provide a systemic-risk model, the adversarial-testing duty is live now.


Article 15 robustness and cybersecurity — what to attack

For a high-risk system, Article 15 requires resilience to errors, faults and inconsistencies, and to attempts by third parties to exploit vulnerabilities and alter the system's use, behaviour or performance. Red-teaming maps onto that directly. Cover at least:

  • Adversarial inputs — crafted inputs designed to force a wrong or unsafe output.
  • Data poisoning — corrupting training or fine-tuning data to bias the model.
  • Model extraction / inversion — stealing the model or reconstructing training data.
  • Prompt injection and jailbreaks — for generative components, bypassing the intended constraints.
  • Robustness under noise and edge cases — degraded, malformed or out-of-distribution input.
  • Feedback-loop manipulation — gaming a system whose outputs influence its future inputs.

Each scenario gets a documented method, a result, and a severity — and findings feed back into the Article 9 risk-management system and the Annex IV technical documentation.


Article 55 — adversarial testing for systemic-risk GPAI

A GPAI model crosses into systemic-risk territory under Article 51 (the threshold uses, among other things, a 10²⁵ FLOP training-compute presumption). For those models, Article 55 layers on extra duties: perform model evaluation including adversarial testing to identify and mitigate systemic risks, assess and mitigate possible systemic risks at Union level, track and report serious incidents, and ensure an adequate level of cybersecurity protection. Red-teaming is the named mechanism for the evaluation-and-mitigation part. The Article 55 duties bind GPAI model providers — a different and much smaller population than high-risk system providers — and fines for GPAI model providers run under Article 101 (Commission-imposed, €15M/3%), not the Article 99 system used for AI systems.

A practical note: most companies reading this are not systemic-GPAI providers. If you fine-tune or deploy someone else's model, you are almost certainly a deployer or a downstream provider, not the model provider — your red-teaming serves Article 15 on your own high-risk system, not Article 55.


The template — sections to capture

1. Target and objective — the system or model under test; whether the goal is Article 15 robustness/cybersecurity evidence or Article 55 systemic-risk evaluation; the assets you are protecting.

2. Scope and rules of engagement — what is in and out of scope; environments; data handling; who is authorised; the non-disruption boundary for any live-system testing.

3. Threat model and scenarios — the attacker profiles and the concrete scenarios (adversarial inputs, poisoning, extraction, jailbreaks, noise/edge cases); the method for each.

4. Severity scale — a defined scale (e.g. critical / high / medium / low) keyed to impact on safety, fundamental rights, security and continued compliance; the criteria for each band.

5. Findings log — each finding: scenario, reproduction, severity, affected component, evidence.

6. Remediation — the fix or mitigation per finding; owner; due date; residual risk after fix; the link into the Article 9 risk-management record and Annex IV docs.

7. Re-test and cadence — verification that fixes hold; the trigger to re-run (model retrain, new capability, major release); reporting line.

8. Incident linkage — when a finding indicates a realised serious incident, the path to Article 73 (high-risk systems) or the Article 55 incident-tracking duty (systemic GPAI).


Worked example: a SaaS provider's document-classification model

Aktenklar GmbH, a provider with about 60 employees and roughly €11M turnover, ships a high-risk AI system that classifies and routes case documents for legal-aid eligibility (Annex III, access to essential services). It is not a GPAI model provider, so its red-team plan targets Article 15, not Article 55. The plan scopes three scenarios: adversarial inputs crafted to force a wrong eligibility class, prompt-injection against the generative summary feature, and robustness under scanned, low-quality documents. Each finding gets a severity on Aktenklar's four-band scale; a critical jailbreak that exposed another applicant's data is fixed, re-tested, and logged into the Annex IV documentation and the Article 9 risk register. Because the jailbreak briefly exposed personal data, Aktenklar assesses it against the serious-incident threshold and builds the Article 73 reporting path it will owe once the high-risk regime applies on 2 December 2027. The red-team plan is the artefact that turns "we tested it" into a dated, reproducible compliance record.


How Confir helps

Confir's data-and-technical-robustness assessment records the red-team scope, scenarios, severity findings and remediation as structured evidence and links each finding to the Article 15 robustness/cybersecurity requirement, the Article 9 risk-management record, and the Annex IV documentation. Where a finding meets the serious-incident threshold it surfaces the Article 73 path. It also helps a provider confirm, by use and role, whether it owes the Article 15 high-risk duty or the Article 55 systemic-GPAI duty. The engine is deterministic and rule-based.


Manage your EU AI Act compliance in one place

Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.

Start free trial →