International AI Security Law

How AI security testing laws differ across jurisdictions: EU AI Act, US executive orders, UK AI Safety Institute, China AI regulations, and their impact on red teaming scope.

AI regulation is evolving rapidly across jurisdictions, and no two countries take the same approach. For red teamers, this means the legality, scope, and reporting obligations of a given engagement can change dramatically depending on where the system is deployed, where the tester is located, and where the data resides.

Regulatory Approaches Compared

European Union: The AI Act

The EU AI Act is the world's first comprehensive AI regulation and has the most direct impact on red teaming activities.

Risk Classification

Red Teaming Obligations Under the AI Act

For high-risk AI systems, Article 9 requires providers to implement a risk management system that includes:

• Identification and analysis of known and foreseeable risks
• Testing with "previously unknown inputs" (adversarial testing)
• Evaluation against "reasonably foreseeable misuse"
• Testing against risks to health, safety, and fundamental rights

General-Purpose AI (GPAI) Models

Models with systemic risk (generally those trained with more than 10^25 FLOPs) face additional obligations:

• Adversarial testing to identify systemic risks
• Documented testing methodology
• Incident reporting for serious incidents
• Cybersecurity protections for model weights

United States: Fragmented Federal Approach

The US lacks a comprehensive federal AI law. Instead, AI testing obligations arise from:

Executive Order 14110 (2023)

• Dual-use foundation model developers must share red team results with the government
• Federal agencies must conduct red team testing before deploying AI
• NIST directed to develop AI testing standards (see NIST AI RMF)

Sector-Specific Regulations

State-Level Activity

States are filling the federal gap with their own AI legislation. Colorado, California, Illinois, and Texas have enacted or proposed laws with testing requirements. See the legal frameworks overview for a state-by-state breakdown.

United Kingdom: Pro-Innovation Framework

The UK has taken a deliberately different approach from the EU, favoring flexibility over prescriptive regulation.

AI Safety Institute (AISI)

The UK AI Safety Institute conducts pre-deployment testing of frontier AI models. Key aspects:

• Voluntary agreements with major AI labs for pre-release access
• Evaluation capabilities for dangerous capabilities, societal harms, and misuse potential
• Published research on evaluation methodology and findings
• International coordination with US AISI and other bodies

Regulatory Sandbox Approach

The UK encourages security research through regulatory sandboxes and the principle that existing regulators (FCA, Ofcom, CMA, etc.) should apply AI governance within their existing mandates rather than creating new AI-specific laws.

China: Content Control Focus

China's AI regulations focus heavily on content control and alignment with "socialist core values." Key instruments:

• Interim Measures for the Management of Generative AI Services (2023): Requires safety assessments before public-facing generative AI services launch
• Algorithm Recommendation Regulation (2022): Requires algorithmic transparency and user control
• Deep Synthesis Regulation (2023): Governs deepfakes and synthetic media

Red Teaming Implications

Navigating Multi-Jurisdictional Engagements

Many AI red teaming engagements involve multiple jurisdictions. A common scenario: a US-based tester evaluating an AI system deployed by a UK company serving EU users, built on a model from a US provider.

1. 1

   Map all jurisdictions

   Identify where the tester is located, where the system is hosted, where the model provider is incorporated, and where the end users are located.

2. 2

   Identify the most restrictive requirements

   The EU AI Act's extraterritorial reach means EU rules often apply even when the system is hosted outside the EU. Apply the most restrictive standard.

3. 3

   Address data transfer rules

   GDPR and other data protection laws may restrict transferring test data (including model outputs) across borders. Ensure your data handling plan accounts for this.

4. 4

   Align on reporting obligations

   Different jurisdictions have different incident and vulnerability reporting requirements. Your engagement contract should specify which reporting obligations apply.

Related Topics

• Legal Frameworks for AI Red Teaming -- US-focused legal analysis including CFAA
• EU AI Act Compliance Testing -- detailed EU AI Act testing methodology
• Authorization, Contracts & Liability -- contract provisions for multi-jurisdictional engagements
• NIST AI RMF & ISO 42001 -- the frameworks regulators reference

References

• "EU Artificial Intelligence Act" - European Parliament (2024) - Articles 9 and 15 on adversarial testing requirements for high-risk AI systems
• "Interim Administrative Measures for Generative AI Services" - Cyberspace Administration of China (2023) - China's regulatory framework for generative AI including security assessment requirements
• "AI Safety Institute Approach to Evaluations" - UK AI Safety Institute (2024) - UK government methodology for pre-deployment AI safety testing
• "Cross-Border Data Transfer Mechanisms and AI Testing" - International Association of Privacy Professionals (2024) - Analysis of GDPR implications for international AI security assessments
• "ISO/IEC 42001:2023 AI Management Systems" - International Organization for Standardization (2023) - International standard for AI governance referenced across jurisdictions