The European Commission published common rules for AI regulatory sandboxes in the EU’s Official Journal on 11 November 2026, setting a shared framework for supervised development and testing. The regulation takes effect on 1 December. For research spinouts and prospective AI providers, it clarifies admission, project plans and exit documentation while preserving responsibility for compliance, according to the adopted regulation.
Common procedures for supervised AI development
Commission Implementing Regulation (EU) 2026/2487, adopted on 10 November, covers the establishment, operation and supervision of all sandboxes created under Article 57 of the EU AI Act. Its stated purpose is to prevent fragmentation between national schemes and make their results mutually recognisable. These are controlled environments for developing, training, validating and testing innovative systems under an agreed plan, including real-world testing where appropriate.
The new rules supply operational detail for an existing AI Act framework. They do not announce that every national sandbox is already open. The Commission’s current Article 57 text, based on the consolidated Act of 27 July 2026, requires at least one national sandbox to be operational by 2 August 2027. Joint arrangements can meet that obligation where they provide equivalent national coverage.
Who can apply and how projects are selected
Under Article 3 of the implementing regulation, providers and prospective providers can apply, including in partnerships with deployers or other third parties needed for the project. Applications can concern a system awaiting market placement or use, an existing system facing substantial modification, or a previously non-high-risk system whose changed use brings it into the high-risk category specified by Article 25(1)(c). Applicants are not limited to their home Member State.
Authorities must consider regulatory and practical challenges, the potential for transferable regulatory learning, public interest or societal benefit and EU policy, proposal maturity, and the system’s innovative character. They may prioritise particular sectors according to resources and capacity while respecting broad and equal access. The selection outcome is due within three months of a complete application or the deadline of a call for applications; applicants have appeal rights under the relevant procedures.
For university spinouts, the relevance is their possible role as providers or prospective providers. Academic affiliation alone does not establish eligibility or guarantee selection. The broader distinction between institutional roles is explained in CASRAI’s background on AI Act obligations and exemptions for research organisations.
The project plan sets the scope of supervision
Article 5 requires the authority and selected participant to agree a sandbox plan. It specifies the estimated timeline, objectives, activities, responsibilities and services, together with the legal requirements being supervised. It must indicate whether personal data is processed and whether the competent data protection authority needs to be involved. Risk safeguards and procedures for serious incidents are required when necessary and appropriate to the risks; a real-world testing plan is included where applicable.
This makes the agreed project scope a concrete record for a university research office: it identifies the responsibilities of the provider, deployers and other participants, and the obligations the authority will examine. The requirement to address personal-data processing also connects sandbox planning to the wider subject of AI data governance. The Commission’s current Article 57(10) requires relevant data protection authorities to be involved within their respective powers.
The regulation explicitly recognises research and experimentation labs in its wider innovation ecosystem. Recital 18 encourages their involvement where appropriate; Article 4 requires cooperation with relevant infrastructure such as testing facilities, European Digital Innovation Hubs and AI Factories. Authorities must securely record sandbox activities and keep records of systems developed or tested there.
Exit documents support assessment without granting approval
Article 6 distinguishes written proof of activities and evaluated requirements from an exit report covering regulatory issues, solutions, lessons and any relevant serious incidents. Market surveillance authorities and notified bodies across the EU must recognise both as documentation of the work and its outcomes. Neither has the legal status or effect of a declaration of conformity. The regulation’s recital 20 also states that participation and exit do not authorise market placement or create a presumption of conformity.
The exit report must reach the participant within a reasonable period and no later than 45 working days after participation ends. For a research commercialisation project, its value therefore depends on the activities and requirements actually examined. It can inform subsequent assessment, but cannot substantiate a broader claim that the system has received regulatory clearance.
Confidentiality is another significant issue for unpublished research and university IP. Article 5 requires participants to identify potentially confidential information in the plan to the best of their ability. Publication of the exit report on the EU interface requires explicit agreement from both parties. Authorities can nevertheless publish non-confidential findings without that agreement, and information not marked confidential in the plan or before project completion is presumed available for disclosure, subject to the applicable access-to-documents rules.
Authorities and reporting emphasise the limits
Greece’s Special Secretariat of Artificial Intelligence and Data Governance said on 13 November that the common framework aims to prevent fragmentation, support consistent implementation and enable mutual recognition of results. Its announcement also stressed that successful completion does not provide market authorisation or a presumption of conformity.
Independent coverage highlighted similar boundaries. TechTimes’ Clark Johnson, reporting on 12 November, focused on the distinction between recognised evidence and compliance clearance, alongside continuing liability. In Cybersecurity360 on 11 November, Paolo Tarsitano emphasised how project plans bring security testing, personal-data questions and confidentiality into supervised experimentation.
What remains and what happens next
Under the Commission’s current Article 57(12), participants remain liable under applicable law for harm to third parties. Protection against administrative fines for AI Act infringements depends on observing the plan and participation conditions and following the authority’s guidance in good faith. Protection concerning other laws requires the relevant authorities’ active involvement and compliance guidance. This is a conditional arrangement, rather than a blanket liability exemption.
The immediate legal milestone is 1 December 2026, when the implementing regulation enters into force, twenty days after Official Journal publication. The national operational deadline remains 2 August 2027. Article 8 also allows authorities to suspend testing or participation where no mitigation is possible, including significant unmitigated risks or breaches of the plan, and requires them to inform the AI Office. For participating research teams, supervision and the agreed safeguards remain central throughout the project.
Sources
- EU Publications Office: Commission Implementing Regulation (EU) 2026/2487, adopted 2026-11-10, Official Journal publication 2026-11-11; full English text and metadata checked 2026-11-16.
- European Commission, AI Act Service Desk: Article 57, consolidated text as at 2026-07-27; checked 2026-11-16.
- Greece, Special Secretariat of Artificial Intelligence and Data Governance: implementing regulation announcement, 2026-11-13; checked 2026-11-16.
- TechTimes, Clark Johnson: EU AI regulatory sandbox rules, 2026-11-12; checked 2026-11-16.
- Cybersecurity360, Paolo Tarsitano: AI Act sandbox rules and security, 2026-11-11; checked 2026-11-16.







