Regulations in the GenAI Era: What Enterprises Need to Know

The regulatory picture in 2026

GenAI regulation is no longer a forward-looking concern. It is a present operational constraint on how enterprises plan, build and ship AI systems.

What has changed since most compliance guides were written is that the landscape stopped being a single wave and became several currents moving at different speeds. The EU tightened its rules, then loosened its timeline. The US pulled back federally while its states pushed forward. China kept legislating quickly and narrowly. Two jurisdictions now have comprehensive horizontal AI laws in force, the EU and South Korea, and roughly 38 US states have enacted AI measures of some kind.

The specific requirements differ, the common thread does not. Every major framework expects you to demonstrate that your AI systems have been tested, monitored and governed across their lifecycle, rather than reviewed once at launch.

Why a unified view matters

Most enterprises are already subject to several AI obligations at once, across jurisdictions, product lines and business units. The hard part is not interpreting any single regulation. It is building one compliance posture that absorbs overlapping requirements and different risk tiers without fragmenting into parallel processes that slow development to a crawl.

A unified view also guards against the most expensive mistake in this area: treating compliance as a project with an end date. Most of these frameworks require continuous evidence, not a one-time submission.

Gaps that surface repeatedly in enterprise assessments:

• No process for detecting when a third-party model update changes system behaviour in ways that affect compliance posture
• Incomplete mapping between what an AI system actually outputs and the obligations that output triggers, particularly where one deployment touches several jurisdictions
• No machine-readable audit trail that can be produced on demand, rather than assembled manually after a request arrives
• Compliance owned by legal alone, which leaves the engineers closest to the risk with the least visibility into the requirements

The frameworks

EU AI Act

The EU AI Act remains the most comprehensive AI law in force. It classifies systems as unacceptable risk (prohibited), high risk (a closed list of uses including hiring, credit scoring, biometrics and critical infrastructure, subject to conformity assessment), limited risk (transparency obligations), or minimal risk.  These are not mutually exclusive: transparency obligations under Article 50 and GPAI obligations under Chapter V can apply cumulatively with high-risk classification to the same system. For GenAI, providers of general purpose AI (GPAI) models must maintain technical documentation, comply with copyright law, publish summaries of training data, and, where a model is designated as posing systemic risk, conduct documented adversarial testing and report serious incidents.

The timeline changed materially in July 2026. Regulation (EU) 2026/1744, the Digital Omnibus on AI, deferred most high-risk obligations to December 2027 for stand-alone systems and August 2028 for AI embedded in regulated products. Other duties did not move. 

Transparency obligations to disclose AI interaction and label AI-generated content took effect in August 2026, alongside the Commission's power to fine GPAI providers. Generative AI systems already on the market before that date have until 2 December 2026 to implement the machine-readable marking obligation specifically.A new prohibition on AI generating non-consensual intimate imagery or child sexual abuse material follows in December 2026, and it reaches further than purpose-built tools: any system where such output is a reasonably foreseeable and reproducible result, and which lacks adequate technical safeguards to reliably prevent it, falls within scope. For enterprises shipping image, video or audio generation, that is an engineering obligation rather than a policy one.

Penalties reach  the higher of €35 million or 7% of global annual turnover for prohibited practices and the higher of €15 million or 3% for most other breaches. Any plan built around the original August 2026 date needs rechecking in both directions.

United States: federal retreat, state advance

The US has no comprehensive federal AI statute, and none appears imminent. Anyone still working from guidance written in 2023 or 2024 should note that the position has inverted twice since then.

The 2023 Executive Order on Safe, Secure, and Trustworthy AI, which established federal safety testing and reporting expectations, was revoked in January 2025. Federal policy now runs through a different set of instruments: America's AI Action Plan of July 2025, an executive order in December 2025 directing agencies to challenge state AI laws considered overly burdensome, legislative recommendations published in March 2026 proposing that Congress preempt such laws, and a June 2026 executive order covering AI-enabled cyber defence and benchmarking for frontier models. Congress has repeatedly declined to enact preemption, so as of mid-2026 no state law has actually been displaced.

The practical consequence is that binding US obligations sit at state level. California's Transparency in Frontier Artificial Intelligence Act took effect in January 2026. Texas, Illinois and Utah have duties in force. New York's RAISE Act was signed in December 2025. Colorado replaced the timeline of its original AI Act, with duties now beginning in January 2027. Until a statute or a court decision says otherwise, the prudent posture is to comply with state law while tracking the preemption fight.

The NIST AI Risk Management Framework remains voluntary but is widely referenced, and its four functions of GOVERN, MAP, MEASURE and MANAGE are a reasonable internal scaffold. Its Generative AI Profile is the more directly useful companion document for GenAI teams, since it maps specific GenAI risks to suggested actions.

China

China operates the most operationally demanding regime currently in force, built from targeted measures rather than one horizontal law. The Interim Measures for Generative AI Services require providers offering GenAI in China to conduct security assessments, obtain approval for certain services, filter content against Chinese legal standards and protect user data. Separate provisions govern algorithmic recommendation and deep synthesis.

The most demanding piece for GenAI builders is content labelling. The Measures for Labelling AI-Generated and Synthetic Content, in force since September 2025, require both explicit labels visible to users and implicit labels embedded in metadata, across text, image, audio and video. An amended Cybersecurity Law referencing AI became enforceable in January 2026, and draft measures covering anthropomorphic interactive AI services and digital virtual humans went through consultation during 2026. A comprehensive national AI Law remains in draft.

For multinationals, these obligations run parallel to Western frameworks and generally cannot be satisfied by the same controls.

United Kingdom

The UK has not enacted a horizontal AI law. It regulates through existing sector regulators, including the FCA, ICO and CQC, applying cross-sector principles of safety, transparency, fairness, accountability and contestability within their domains. Supporting material such as the AI Cyber Security Code of Practice carries practical weight without being a statute, and government policy has leaned toward statutory regulatory sandboxes rather than new AI-specific duties. Various AI bills have been introduced without becoming law. For enterprises in regulated UK sectors, the operative obligation is your existing regulator's expectations, not a general AI compliance programme.

South Korea

South Korea's AI Framework Act took effect in January 2026, making it the second jurisdiction with a comprehensive horizontal AI law. It is worth tracking as a signal, since it broadly follows the EU's risk-based structure and suggests where other jurisdictions may land.

Standards: ISO 42001, NIST, MITRE ATLAS, OWASP

Standards are increasingly how enterprises demonstrate governance across jurisdictions rather than for any single one. ISO/IEC 42001 is a certifiable AI management system standard. MITRE ATLAS catalogues adversarial tactics and techniques targeting AI systems and is directly useful for threat modelling and red teaming, as is the OWASP Top 10 for LLM Applications. Aligning to these produces documentation and risk practices that travel between regulatory contexts, which is why they are worth the effort even though none of them is law.

What every framework asks for

1. Risk classification of your AI systems
2. Technical documentation and record-keeping
3. Transparency toward users and regulators
4. Human oversight for consequential applications
5. Incident detection, reporting and response
6. Ongoing monitoring rather than launch-time review

Build your infrastructure around those six and you can adapt to individual regulations as they land, instead of rebuilding for each one.

Where enterprises fall short

Four gaps recur.

1. Documentation that does not scale. Frameworks like the EU AI Act expect detailed records covering training data, architecture, intended use and known limitations. Many enterprises can produce this for one flagship system and nothing else.
2. Red teaming that is really quality assurance. Adversarial testing is explicitly required for GPAI models with systemic risk under the EU AI Act, and it is embedded in most other frameworks as an expectation. Internal QA checks whether a system works as intended. Red teaming checks how it fails when someone is deliberately trying to break it. Teams relying on the former are underestimating their exposure, usually significantly. This is the gap Alice was built to close.
3. Fragmented governance. Requirements that span data provenance, model risk and incident escalation cross engineering, security, legal and product. Where those functions coordinate loosely, obligations fall between them.
4. Monitoring treated as optional. Continuous evaluation of behaviour in production, including drift, output degradation and policy misalignment, is now an expectation rather than a maturity signal. It is also the requirement most often deferred, because it is the one that never finishes.

Practical steps

1. Inventory your AI systems. Every system in use or development. For each: jurisdiction, use case, risk tier, applicable frameworks, and an owner.
2. Build documentation infrastructure. Processes that capture data provenance, model documentation, test results and deployment records as work happens, rather than reconstructing them under deadline.
3. Integrate adversarial testing. Structured red teaming before deployment and continuously in production, informed by real-world attack intelligence so tests reflect actual adversary behaviour rather than a checklist.
4. Establish cross-functional ownership. Named accountability across legal, compliance, security and engineering, with defined escalation paths for incidents and policy questions.
5. Deploy continuous monitoring. Observability that tracks behaviour in production, detects drift and regressions, and produces audit-ready logs on demand.

Steps 1, 2 and 4 are governance work, and you will need people and process for them. Steps 3 and 5 are engineering work, and that is where Alice comes in.

How Alice supports regulatory compliance

Alice handles the part of compliance that has to be proven rather than promised: producing and keeping the evidence that your safeguards actually work.

It runs as one connected loop rather than three separate checks, which matters because the frameworks increasingly ask for exactly that.

Red teaming finds how your AI breaks. 

Automated attacks probe for prompt injection, model manipulation, data exfiltration and agents acting outside their intended scope, aligned to reference sets including MITRE ATLAS and OWASP. You get a full assessment: every attack run, whether it passed or failed, why, and how to fix it. That is the documented adversarial testing these frameworks call for, in a form your security and legal teams can sign off on.

Those findings become your guardrails. 

Rather than bolding on generic safeguards, Alice's guardrails are formulated from your own red teaming results and trained on your exact policies, then run inline at low latency, P95 ~120ms, so they block real risks instead of burying your team in false alarms. Because they are policy-driven, they can be configured for jurisdiction-specific content requirements, which is what makes multi-jurisdiction deployment tractable.

Continuous testing keeps both sharp. 

Red teaming keeps probing for new failure modes as models update and attacks evolve. Every finding updates your guardrails, and what those guardrails see in production sharpens the next round of tests. That closes the monitoring gap, and it is close to a literal description of what regulators mean by ongoing rather than launch-time review.

Throughout, you get a record. 

A trail showing what happened, when, and what your safeguards did about it, which is what you hand a regulator, an auditor, or an enterprise customer's security team.

Underneath it sits Rabbit Hole, our threat intelligence engine: more than a decade of knowledge about how malicious actors actually think and operate, now pointed at AI. It supplies the real-world attacks used to test your system, takes known attacks apart to generate new variations, and updates as new tactics appear in the wild.

We spent a decade defending the platforms billions of people use every day against the most sophisticated bad actors online. That is why the leading foundation model labs trust us with their models, and why enterprises building on those models can get the same protection.

What we do not do

We do not write your technical documentation, run your conformity assessment, build your governance structure or manage your regulatory filings. Those are legal and process workstreams.

What we do is generate and maintain the evidence all of that paperwork is supposed to point at, and make your AI genuinely safer in the process.

Need help preparing for the next era of AI regulation? Talk to Alice's team.

‍