A real-time control layer for the obligations that are already in force.

Banned uses and high-risk classification

Article 5 prohibits specific AI practices outright — including real-time remote biometric categorization in public spaces for law enforcement (with narrow exceptions), AI systems that exploit vulnerabilities to manipulate behavior, and social scoring systems. These prohibitions are not subject to the full compliance timeline; they applied from February 2025. Article 6 defines high-risk AI systems by reference to Annex III, which covers systems used in critical infrastructure, employment, education, essential services, law enforcement, migration, and justice. Deployers of high-risk systems carry independent obligations regardless of whether they built the system.

• Assessment of whether any deployed AI system falls within a prohibited practice under Article 5
• Classification of AI systems against Annex III high-risk categories
• For high-risk systems: conformity assessment, technical documentation, and registration
• Deployer obligations under Article 26: monitoring, human oversight, and incident reporting

Ongoing risk management and operational control

Article 9 requires providers of high-risk AI systems to implement a risk management system that operates throughout the entire lifecycle of the AI system — including post-deployment. Article 26 places obligations on deployers to implement appropriate technical and organizational measures for human oversight, to monitor operation, and to report serious incidents. The obligation is operational, not documentary: a risk management system that exists only in policy documents, with no real-time control on what the AI system actually does, does not satisfy the Article 9 standard.

• Risk management system covering the full AI system lifecycle, including deployment
• Technical measures enabling human oversight of high-risk AI system outputs
• Monitoring of AI system operation for deviations from expected behavior
• Incident reporting procedures for serious incidents involving high-risk AI systems

General-purpose AI model governance and acceptable-use enforcement

Articles 53 through 55 impose obligations on providers of general-purpose AI models with systemic risk, including adversarial testing, incident reporting to the AI Office, and cybersecurity measures. Organizations that deploy GPAI models in agentic configurations — where the model autonomously selects tools, invokes APIs, and takes actions — inherit exposure under both the GPAI provisions and the high-risk deployer obligations. The maximum fine for GPAI violations is 7% of global annual turnover.

• Assessment of whether deployed GPAI models carry systemic risk under Article 51
• Acceptable-use policy enforcement for GPAI model integrations
• Adversarial testing documentation where required
• Real-time monitoring of GPAI model behavior in agentic configurations

An AI agent in a customer service context generates a response that uses a persuasion technique targeting a user's demonstrated cognitive vulnerability — a prohibited practice under Article 5(1)(b). Compiled intercepts the response before it is delivered, blocks it, and logs the finding with attribution to Article 5 of the EU AI Act.

An AI system used for employment screening — a high-risk category under Annex III — generates a recommendation that would automatically advance or reject a candidate without human review. Compiled flags the output before it is acted on, routing a finding to the designated human reviewer under the Article 26 oversight obligation.

An AI agent uses a GPAI model integration to perform a task that falls outside the organization's defined acceptable-use policy for that model — generating content that violates the provider's usage terms in a way that creates regulatory exposure. Compiled blocks the invocation before the output is generated and logs the attempted use with policy attribution.

A GPAI-powered agent begins autonomously invoking additional APIs and accumulating permissions not defined in its original scope — a pattern the OWASP Top 10 for Agentic Applications (Dec 2025) identifies under privilege escalation. Compiled detects the scope expansion and flags it for human review before the agent gains additional access.

Every AI-agent action is assessed before it completes. There is no gap between the policy and the enforcement. A prohibited practice is blocked before it affects a user; a high-risk output is flagged before it is acted on.

Compiled's continuous, real-time assessment of agent behavior constitutes the operational risk management system Article 9 requires. The audit trail demonstrates that the system "operates on a continuous basis" as the regulation specifies.

Flagged outputs are routed to a human reviewer before being acted on — the technical implementation of the deployer's Article 26 oversight obligation. The reviewer's decision is logged in your environment.

Antibodies tuned to Article 5 prohibited practices detect manipulation techniques, social scoring patterns, and other banned behaviors in real time — before a user is affected and before the violation is complete.

Your GPAI model acceptable-use policies and scope restrictions are compiled into antibodies. An agent action outside the defined scope is blocked or flagged, with the specific policy cited in the finding.

Every decision is timestamped, scored, and policy-attributed inside your environment. The record demonstrating that your Article 9 risk management system was operating is generated automatically, in your own infrastructure.