In July 2026, the EU successively released an Artificial Intelligence and Cybersecurity Action Plan and Guidelines on AI Transparency Obligations, and refined the supporting mechanisms for general-purpose artificial intelligence model regulation as it enters the formal enforcement phase. Together, these several moves show that AI safety governance is shifting from principle-setting toward engineered execution. Model evaluation, content labeling, operational logging, and incident reporting are beginning to become inspectable and traceable objects of regulation, and will further influence the allocation of responsibility, product delivery, and market competition across the AI industry chain.
EU AI Safety Regulation Shifts from Rule-Setting to Implementation
In July 2026, the EU advanced three sets of key actions around AI safety. The Cybersecurity and Artificial Intelligence Action Plan, released on July 7, brings advanced AI models into the European cybersecurity governance framework. This plan links up with the AI Act, the NIS2 Directive, the Cyber Resilience Act, DORA, and the Cyber Solidarity Act, focusing on building model security assessment capabilities and an AI cybersecurity testing platform, and on promoting the sharing of risk information and testing resources among member states, enterprises, and research institutions.
This plan extends the governance focus of AI safety from content risk to system risk. As models come to possess capabilities such as code generation, vulnerability discovery, tool invocation, and continuous execution, AI can both improve the efficiency of threat detection and vulnerability remediation, and lower the threshold for automated attacks. Regulation therefore no longer merely judges whether a model’s output is illegal or harmful, but also has to assess which systems the model can access, which operations it can execute, and whether risks might be transmitted to enterprise networks and critical infrastructure.
The Article 50 transparency obligation guidelines of the AI Act, released on July 20, translate content transparency into executable requirements. The guidelines clarify the responsibilities for disclosing identity when humans interact with AI, machine-readable marking of synthetic content, deepfake disclosure, and the labeling of some public-interest content. Model providers need to ensure that synthetic content is technically identifiable, while deployers and content publishers need to complete disclosure according to the use scenario. The difficulty of governance shifts from whether a label is added, to whether the marking can be continuously retained after editing, transcoding, and cross-platform dissemination.
Regulation of general-purpose artificial intelligence models also entered the final stage of enforcement preparation in July. The relevant obligations already began to apply in August 2025: general-purpose model providers need to prepare technical documentation, copyright policies, and summaries of training content; models with systemic risk additionally need to conduct risk assessments, implement mitigation measures, report major incidents, and strengthen cybersecurity protection. From August 2, 2026, the European Commission began to formally exercise its powers of inspection and fines.
Together, the three sets of actions show that the focus of EU AI safety regulation has shifted from setting principles to establishing an evidence system. Model evaluation, version records, operational logs, content labeling, and incident reporting will become the basis for regulatory inspection and the determination of responsibility. AI safety is no longer a one-time review before launch, but a continuous governance requirement covering development, deployment, operation, and upgrading.
AI Safety Responsibility Will Be Reallocated According to Information, Decision-Making, and Control Capabilities
Generative AI products are usually composed of foundation models, fine-tuned models, knowledge bases, enterprise data, plugins, and business systems together. Risk often arises from the combination of multiple links: the foundation model determines the boundary of capabilities, the application design determines the data and tools the model can invoke, and the deploying enterprise determines which business processes the system enters. The traditional division of responsibility centered on a single supplier is hard-pressed to cover this multi-layered structure.
Foundation model vendors master the training methods, capability boundaries, known defects, and evaluation results, and therefore need to bear model-level risk identification, the provision of technical materials, the disclosure of major risks, and the construction of security interfaces. Their responsibility is not to backstop all downstream applications, but to provide sufficient information downstream so that application vendors can judge whether a model is suitable for a specific use. When a model upgrade adds capabilities such as code generation, tool invocation, or autonomous planning, the change in risk also needs to be explained in tandem. The EU’s current rules already require general-purpose model providers to provide necessary technical information downstream, and to fulfill stricter assessment and reporting obligations for models with systemic risk.
Application vendors determine how general-purpose capabilities are transformed into specific products, and bear the responsibility for scenario risk control. The same model used for document summarization mainly carries the risk of content distortion; used for recruitment, credit, healthcare, or production control, errors may directly affect individual rights and business operations. After an agent connects to email, databases, or payment systems, the security focus also shifts from answer accuracy to permission boundaries, operation confirmation, abnormal termination, and human takeover.
Deploying enterprises determine which processes AI enters and whether outputs directly form decisions, and therefore still need to bear responsibility for the scope of use, human review, continuous monitoring, and internal accountability. Content platforms, meanwhile, are responsible for source identification, label retention, and necessary disclosure on the dissemination end. In the future, responsibility will not simply transfer from upstream to downstream, but will be configured in layers according to who masters the risk information, who decides the use scenario, and who is able to take control measures.
Industry Players Will Reconstruct Product Delivery, Procurement, and Service Models
Foundation model vendors need to incorporate safety materials into standard delivery. In addition to comparing model performance, price, and latency, enterprise customers will also demand capability boundaries, testing methods, risk records, version changes, and incident response plans. What model vendors deliver will not only be APIs, but also documentation, interfaces, and tools that support customers in carrying out risk judgment and continuous governance.
Application and agent vendors need to write permission control into the product architecture. Least privilege, secondary confirmation of sensitive operations, complete execution logs, abnormal circuit-breaking, and human takeover will become basic conditions for agents to enter enterprise systems. Different scenarios also require different levels of control strength; content generation, customer service, personnel screening, and production control cannot use the same set of security configurations.
Enterprise procurement will add stages of security admission and continuous evaluation. Beyond verifying effectiveness and return on investment, AI pilots also need to check whether the system is auditable, whether model updates are traceable, whether anomalies can be located in a timely manner, and whether suppliers can continuously provide security updates. In fields such as finance, healthcare, energy, and public services, security evaluation and incident response capabilities may be written directly into tender conditions and contract terms.
Content platforms need to build cross-platform source governance capabilities. Machine-readable marking, digital watermarking, content credentials, and metadata standards will gradually become platform infrastructure, and the core difficulty is whether the markings of different models can be mutually recognized, and whether content can continue to be traced after editing and reposting. Cybersecurity and professional service institutions, meanwhile, may extend into model red-team testing, agent permission governance, operational monitoring, and AI supply chain auditing.
AI Safety Investment May Form a New Market Access Threshold
In the short term, model evaluation, content labeling, continuous monitoring, technical documentation, and incident response will increase fixed costs. Large vendors can spread the investment across more products and customers, while small and medium-sized vendors may rely on model platforms, cloud vendors, and third-party tools to obtain safety capabilities. This will strengthen the service boundaries of upstream platforms, and may also increase the technical dependence of downstream enterprises.
In the government, finance, healthcare, energy, and large-enterprise markets, safety investment may in turn transform into an admission capability. Suppliers that can provide complete evaluation materials, accept third-party testing, track model versions, and bear continuous response responsibility can more easily reduce customers’ concerns about business interruption, legal liability, and reputational loss. Model performance determines whether a product can complete a task, while the safety system determines whether a product can enter critical processes.
The future market structure depends on whether safety capabilities can be standardized. If enterprises need to separately build evaluation, labeling, and compliance systems, fixed costs will strengthen the advantage of leading vendors; if third-party testing, industry standards, and general governance tools gradually mature, small and medium-sized enterprises can procure safety capabilities on demand, and model evaluation, content provenance, audit certification, and risk insurance will also form a new professional service market.
The long-term impact of the EU’s July policy changes lies in AI safety beginning to enter the distribution of industrial value. Safety capabilities will simultaneously influence product design, supplier admission, contractual responsibility, and customer trust, and will become a basic condition for AI to enter critical industries and core business processes.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.