On 9 October 2026 the European Commission held a special meeting of the scientific panel of independent experts provided for by the AI Act, with Executive Vice-President Henna Virkkunen announced as attending. According to the press release, the panel of 60 experts has been investigating recent loss-of-control incidents and, together with the AI Office, has worked on a set of questions for the companies that developed the models involved; during the meeting it was to present recommendations to the Commission on frontier AI safety and security risks (European Commission). The press release names neither models nor companies, and to date no outcomes or recommendations have been published (European Commission). What can be examined is the legal framework, whose enforcement powers vis-à-vis model providers have been exercisable since 2 August 2026. An episode of loss of control does not find the AI Act unprepared: it finds a procedure already written, with defined actors, deadlines and powers.
Loss of control in the AI Act: a systemic risk of general-purpose models
Regulation (EU) 2024/1689 does not use the expression “loss of control”: recital 110 merely refers to “unintended issues of control relating to alignment with human intent” among the risks identified by international approaches. The relevant notion is the “systemic risk” of Article 3, point (65): a risk specific to the high-impact capabilities of general-purpose AI models, with negative effects on public health, safety, fundamental rights or society, “that can be propagated at scale across the value chain” (EUR-Lex). The Commission expressly includes loss of control among the risks of large-scale harm that providers of the most advanced models must address, alongside, among others, cyber offence (European Commission).
The operational definition is found in the “Safety and Security” chapter of the General-Purpose AI Code of Practice, which includes it among the specified systemic risks that signatories must take into account: risks from humans losing the ability to “reliably direct, modify, or shut down a model” (European Commission). The link with the enacting terms runs through the “serious incident” of Article 3, point (49), which however refers to an AI system and to its consequences: death or serious harm to health, serious and irreversible disruption of critical infrastructure, infringement of obligations protecting fundamental rights, serious harm to property or the environment. Not every episode of loss of control is, for that reason alone, a serious incident: in our view, it becomes one when it leads, directly or indirectly, to one of those consequences.
The scientific panel and the AI Office: from qualified alert to action
Article 68 requires the Commission to establish, by means of an implementing act, a scientific panel of independent experts “intended to support the enforcement activities under this Regulation”. The act is Commission Implementing Regulation (EU) 2025/454, which caps the panel at 60 members, appointed for two years with the possibility of renewal, independent of any provider of AI systems or models and bound to publish declarations of interests (EUR-Lex). The main tool is the “qualified alert” under Article 90, available where the panel has reason to suspect that a model poses “concrete identifiable risk at Union level” or meets the conditions of Article 51. A decision by at least a simple majority of the members is required, and the AI Office must handle the complete alert within two weeks, either initiating the measures under Articles 91 to 93 or closing it; with the authorisation of at least one third of the members, the panel may also ask the Commission to send a request for information to the provider (Article 91(3)). The press release of 9 October does not state on which basis the questions to the companies were prepared. The qualified alert is a technical alert that obliges the Office to decide, not a finding of infringement.
Article 88 gives the Commission exclusive powers to supervise and enforce the obligations of model providers, entrusted to the AI Office. The Commission may request the documentation under Articles 53 and 55 (Article 91); the Office may evaluate the model “to investigate systemic risks at Union level”, including with experts from the panel (Article 92); the Commission may then request mitigation measures where the evaluation has given rise to “serious and substantiated concern” of a systemic risk, or request that the model’s making available on the market be restricted, or that it be withdrawn or recalled (Article 93). The provider’s commitments may be made binding, and infringements and non-compliance are subject to fines of up to 3 per cent of total worldwide annual turnover or EUR 15 million, whichever is higher (Article 101).
The Digital Omnibus on AI, in force since 27 July 2026, has added a second track. The new Article 75 gives the AI Office exclusive competence over AI systems based on general-purpose models where the model and the system are developed by the same provider or by providers belonging to the same undertaking; Articles 75a to 75d confer on it the powers of a market surveillance authority, the opening of investigations, binding commitments and the non-compliance decision (EUR-Lex). As NicFab observes, this is a procedural model borrowed from competition law, while deployers outside the provider’s undertaking remain subject to national supervision (NicFab), which in Italy lies with the authorities designated by Legislative Decree 179/2026. It follows that the model is governed by Section 5 of Chapter IX, the system built on top of it by Article 75 et seq.
Models with systemic risk: evaluation, mitigation and serious incident reporting
Article 55(1) requires these providers to perform model evaluation in accordance with protocols reflecting the state of the art, including adversarial testing; to assess and mitigate possible systemic risks and their sources; to keep track of, document and report “without undue delay” to the AI Office relevant information about serious incidents and possible corrective measures; and to ensure adequate cybersecurity of the model and of its physical infrastructure. Paragraph 2 allows compliance to be demonstrated by adhering to a code of practice.
The “Safety and Security” chapter, which the Commission and the AI Board have confirmed as an adequate voluntary tool (European Commission), sets that obligation against a timetable: an initial report within two days in the case of serious and irreversible disruption of critical infrastructure, within five in the case of a serious cybersecurity breach, including the (self-)exfiltration of model weights, within ten in the case of a death, within fifteen in the other cases of serious incident; then intermediate reports at least every four weeks and a final report within 60 days of the incident being resolved. Reporting, the Code specifies, is not an admission of wrongdoing. The Code does not add obligations to the Regulation, but it makes the way they are fulfilled verifiable; those who do not adhere to it must demonstrate alternative adequate means of compliance.
Downstream providers and deployers: what changes for those integrating the models
The AI Act defines as a “downstream provider” whoever integrates an AI model, their own or a third party’s, into their own system (Article 3, point (68)). Article 53(1)(b) requires the model provider to make available information that enables providers of AI systems “to have a good understanding of the capabilities and limitations” of the model. The Code commits signatories to facilitating the reporting of serious incidents by downstream providers, users and third parties, by informing them of the direct reporting channels available. Article 89(2) grants downstream providers the right to lodge a complaint, for which the Commission has set up a dedicated channel. Whoever then puts the system to a high-risk use takes on the obligations of the provider of a high-risk system, as we examined with regard to attributability and liability in the AI value chain.
For deployers, the monitoring and serious-incident reporting obligations of Article 26(5) concern high-risk systems, whose rules apply largely from 2 December 2027 (European Commission). We consider, however, that the interest in learning of an incident affecting the underlying model does not wait for that date: contractual clauses on information flows, model versions and suspension of the service are today the most concrete safeguard. In the same direction goes NicFab’s indication that, where the automated sequence is faster than the possibility of intervention, human control must be placed before the relevant action (NicFab). In the financial sector, supervisory authorities have already asked boards to assess their exposure to the cyber risks of frontier models, as recalled with regard to frontier artificial intelligence in the boardroom.
Critical aspects: transparency, definitions and the timing of the procedure
The first aspect concerns transparency: the implementing regulation provides for the publication of the panel’s opinions and recommendations, except where this would disclose confidential business information, trade secrets or strategic interests of the Union. The exception is understandable, but if models and companies remain anonymous confidentiality risks depriving downstream providers of precisely the information they would need.
The second aspect is definitional: loss of control is defined by a voluntary code and by a recital, not by the enacting terms, and the serious incident is defined for systems, whereas the reporting obligation of Article 55 falls on models. We consider that the two fit together, but the connection is left to the practice of the AI Office.
The third aspect concerns timing: the Code measures reporting in days, the chain from alert to measure is longer, and Article 93 makes mitigation measures conditional on a “serious and substantiated concern”. It is a safeguard for providers that makes timeliness a matter of resources rather than of powers. The institutional sustainability of the supervisory apparatus is also addressed in the analysis by NicFab. Finally, the press release does not warrant conclusions on the seriousness of the episodes: it speaks of investigation, not of findings.
Conclusions
Providers of models with systemic risk must assess and mitigate loss of control, document incidents and report them to the AI Office within the time limits that the Code makes verifiable; the scientific panel is responsible for the alert, the AI Office for the decision on measures. Still missing, for now, are the outcomes of the meeting of 9 October and a legal definition of loss of control. Those who integrate these models would do well to obtain the documentation under Article 53, agree channels and time limits for information on incidents, and place human control before irreversible actions: the AI Act distributes risk management along the value chain, and each link answers for its own part.
Author: Valentina Grazia Sapuppo
