New Yorker RAISE-Gesetz: Sicherheitslücken im Bereich der künstlichen Intelligenz

Publishing an AI safety policy is one thing. Proving that the safeguards behind it actually work is another.

As frontier AI models become more capable, organizations face new questions about what those systems can do, how they could be misused, and whether security controls can prevent catastrophic outcomes.

What happens if a powerful model helps enable a large-scale cyberattack? Can a developer identify dangerous capabilities before deployment? Who investigates when a model circumvents safeguards? And how quickly can the organization report a critical safety incident?

These are the questions at the center of New York’s Responsible AI Safety and Education (RAISE) Act.

Signed into law in December 2025 and substantially amended in March 2026, the RAISE Act establishes transparency, safety framework, and incident reporting requirements for qualifying frontier AI developers. The amended framework is scheduled to take effect 1. Januar 2027.

But the broader lesson extends beyond the developers directly regulated by the law.

AI safety depends on more than documenting risks. Organizations need visibility into the models, data, identities, access pathways, and controls that determine whether those risks can become real-world incidents.

This guide explains the RAISE Act’s requirements, who must comply, what constitutes catastrophic risk, and the governance gaps organizations should address as frontier AI capabilities continue to evolve.

What Is the New York RAISE Act?

Die Responsible AI Safety and Education Act, commonly called the RAISE Act, is a New York law designed to improve transparency and accountability surrounding powerful frontier artificial intelligence models.

Ursprünglich erlassen durch Senate Bill S6953B, the legislation established a framework for addressing severe risks associated with frontier AI development and deployment.

New York subsequently enacted S8828, a chapter amendment that significantly revised the law.

The amended framework aligns several core concepts more closely with California’s frontier AI safety legislation while establishing New York-specific oversight and reporting mechanisms.

Its central focus is catastrophic risk: the possibility that a frontier AI model could contribute to exceptionally severe harm.

The law requires qualifying developers to make specified information about their safety practices available, assess dangerous capabilities, address identified risks, and report critical safety incidents.

Importantly, the RAISE Act does not establish the same obligations for every AI application, enterprise chatbot, or organization deploying a third-party AI service.

Its direct requirements depend on whether an organization meets the law’s definitions of a frontier developer or large frontier developer and whether the relevant model falls within New York’s statutory scope.

When Does the RAISE Act Take Effect?

The amended RAISE Act is scheduled to take effect 1. Januar 2027.

Understanding the legislative timeline matters because older articles may describe requirements from the original version that were changed by the 2026 amendments.

Who Must Comply With the RAISE Act?

The law establishes two important categories: frontier developers und large frontier developers.

These categories are related, but their obligations are not identical.

Under the amended framework, a frontier model is a foundation model trained using more than 10²⁶ integer or floating-point operations, including qualifying subsequent training and modifications.

A large frontier developer is a frontier developer whose annual gross revenues, together with its affiliates, exceeded $500 million in the preceding calendar year.

The law applies to frontier models developed, deployed, or operating in whole or in part in New York State, subject to statutory exceptions.

What Is Catastrophic Risk Under the RAISE Act?

The RAISE Act focuses on severe outcomes that go beyond ordinary model errors, inaccurate responses, or routine AI governance concerns.

Under the amended law, catastrophic risk involves a foreseeable and material risk that a frontier model could cause or materially contribute to an incident resulting in more than 50 deaths or serious injuries, or more than $1 billion in property damage, subject to the statute’s definitions and conditions.

The law is particularly concerned with dangerous capabilities that could contribute to large-scale physical or economic harm.

Potential risk scenarios include:

• Severe cyber threats: Frontier AI capabilities that materially enable exceptionally damaging cyber operations.
• Biological or chemical threats: Model capabilities that could significantly facilitate dangerous misuse.
• Loss of control: Model behavior that circumvents oversight mechanisms in ways that materially increase catastrophic risk.
• Other catastrophic incidents: Qualifying events that meet the law’s severity and causation criteria.

These examples do not mean that every cybersecurity incident, unsafe response, or AI policy violation constitutes a catastrophic risk under the statute.

The legal threshold is intentionally high. The operational challenge is determining whether a model’s capabilities, environment, and safeguards create a credible pathway to that level of harm.

What Does the RAISE Act Require?

The amended law establishes requirements involving safety frameworks, model transparency, incident reporting, and oversight.

1. Published Frontier AI Safety Frameworks

Large frontier developers must write, implement, follow, and publicly publish a frontier AI framework describing how they manage catastrophic risks.

The framework must address specified areas, including:

• Applicable standards and industry best practices.
• Thresholds for identifying dangerous model capabilities.
• Assessments and mitigations of catastrophic risks.
• Safety reviews before deployment or extensive internal use.
• Use of third parties to evaluate risks and mitigations.
• Framework updates and material modifications.
• Cybersecurity practices protecting unreleased model weights.
• Identification and response to critical safety incidents.
• Internal governance and accountability.
• Catastrophic risks associated with internal model use.

Large frontier developers must review their frameworks at least annually and publish material modifications with an explanation within the applicable statutory period.

2. Frontier Model Transparency Reports

Before or concurrently with deploying a new frontier model or substantially modified version, frontier developers must publish specified information about the model.

That information includes the developer’s website, contact mechanism, model release date, supported languages and output modalities, intended uses, and generally applicable restrictions.

Large frontier developers must provide additional summaries concerning catastrophic-risk assessments, their results, third-party evaluator involvement, and related safety steps.

3. Critical Safety Incident Reporting

Frontier developers must report qualifying critical safety incidents to the designated state office within 72 hours after determining an incident occurred or learning facts sufficient to establish a reasonable belief that it occurred.

When an incident poses an imminent risk of death or serious physical injury, the law establishes a separate 24-hour disclosure requirement to an appropriate authority, subject to the statutory conditions.

These requirements make incident identification, escalation, documentation, and internal communication particularly important.

4. Oversight and Regulatory Disclosures

The law establishes oversight responsibilities within the New York Department of Financial Services.

It also includes requirements for large frontier developers to file specified disclosures with the designated office, identify relevant organizational information, and comply with applicable assessments and renewal requirements.

These provisions are distinct from the public-facing frontier AI framework and model transparency reports.

Why Published Safety Policies Aren’t Enough

An organization may have a comprehensive AI safety framework and still struggle to demonstrate that its controls are operating effectively.

Consider a developer that documents model access restrictions, safety evaluations, and incident response procedures.

The framework may look complete. But important questions remain:

• Who can access sensitive training data and unreleased model assets?
• Which service accounts and Maschinenidentitäten have privileged permissions?
• Can unauthorized data movement or access changes be detected?
• Are model security findings connected to remediation owners?
• Can teams investigate a potential incident quickly enough to support reporting?
• Does governance evidence reflect the current environment or a past assessment?

The distinction is significant.

A policy describes what should happen. Operational evidence helps establish what is actually happening.

That is where AI security, data discovery, identity context, and continuous risk visibility become important supporting capabilities.

How an AI Security Incident Exposes the Gap

Imagine a large frontier developer preparing to release a new model.

The organization has published its safety framework, completed model evaluations, and documented relevant mitigations.

During a routine review, the security team discovers that a service identity associated with a development environment has broader access than expected.

The identity can reach sensitive repositories and model-related assets that should be restricted.

There is no immediate evidence of a catastrophic safety incident. But the discovery raises several important questions:

• Which systems and datasets can the identity access?
• Could the permissions expose sensitive information or unreleased model assets?
• When did the access change occur?
• Which teams own the affected resources?
• Has any unauthorized activity occurred?
• What remediation is necessary?
• Does the finding affect the organization’s safety or security assessments?

The organization must investigate before it can determine the significance of the finding.

This example illustrates an important distinction: a security weakness is not automatically a critical safety incident under the RAISE Act. But weak visibility can make it harder to identify, assess, and respond to risks that may escalate.

For developers operating powerful models, that distinction makes reliable security context essential.

What Are the RAISE Act’s Reporting Deadlines?

The amended law establishes distinct reporting requirements for qualifying critical safety incidents.

These requirements should not be confused with ordinary data breach notification rules. The RAISE Act’s critical safety incident provisions address specific frontier AI risks and defined reporting triggers.

Organizations should coordinate legal, security, engineering, and safety teams when establishing response processes.

What Are the Penalties for Violating the RAISE Act?

The amended RAISE Act authorizes the New York Attorney General to bring civil actions involving specified violations.

Potential penalties include:

• Up to $1 million for a first qualifying violation.
• Up to $3 million for each subsequent qualifying violation.

These penalties apply under specified statutory conditions, including certain failures involving required documents, incident reporting, prohibited statements, or compliance with a large frontier developer’s own framework.

Additional provisions address required large frontier developer disclosures, including potential daily penalties for specified filing failures.

The law does not create a new private right of action for violations of the RAISE Act itself.

The enforcement framework reinforces a central point: publishing a safety framework creates responsibilities around following it and accurately representing its implementation.

How Should Organizations Prepare for the RAISE Act?

Direct compliance planning should begin with a determination of whether the organization qualifies as a frontier developer or large frontier developer.

From there, organizations can connect legal requirements with the operational controls needed to support safety, security, transparency, and reporting.

How BigID Supports AI Security and Governance

The RAISE Act’s direct obligations focus on frontier AI safety and catastrophic risk. BigID addresses an important supporting layer: enterprise visibility into AI assets, sensitive data, access relationships, and security exposure.

BigID helps organizations connect AI systems with the data and identities surrounding them, enabling teams to identify risks that may otherwise remain hidden across fragmented environments.

• KI-Asset-Erkennung: Identify AI applications, models, agents, and shadow AI across enterprise environments.
• Ermittlung sensibler Daten: Find and classify sensitive and regulated information associated with AI workflows.
• KI-Zugriffssteuerung: Investigate human and non-human access to sensitive data and connected systems.
• Datenherkunft: Understand data origins, relationships, and movement across supported environments.
• Risk prioritization: Connect sensitivity, access, policy, and business context to identify actionable exposure.
• Nachweise zur Regierungsführung: Support assessments, ownership, remediation, and oversight with relevant data security intelligence.

These capabilities can complement specialized frontier model evaluations, model-weight security controls, and incident reporting systems.

BigID does not independently certify compliance with the RAISE Act or replace the catastrophic-risk assessments and reporting functions required of covered developers.

Its value is helping organizations understand and govern the enterprise data and access environment surrounding their AI systems.

Why the RAISE Act Matters Beyond Frontier Developers

Most organizations will not meet the RAISE Act’s frontier developer thresholds.

But many enterprises are adopting increasingly capable models through cloud services, APIs, embedded applications, and autonomous AI agents.

Those organizations may not face the same direct RAISE Act obligations, yet they still need to answer important security and governance questions.

What sensitive information can an AI system access? Which Identitäten control it? Can the system interact with other enterprise applications? How are risky permissions identified? What happens when an integration changes?

These questions become more important as AI moves from generating content to taking actions across business environments.

New York’s RAISE Act also sits alongside other state approaches to AI regulation. California has adopted frontier AI safety legislation, while other states focus on automated decisions, consumer disclosures, or AI content transparency.

Organizations should avoid treating these different laws as interchangeable.

A sustainable AI governance program establishes a shared foundation for AI visibility, data security, access, and risk, then maps specific legal obligations to the relevant systems and activities.

Entdecken Sie BigIDs Globaler KI-Regulierungs-Tracker to follow regulatory developments across U.S. states and international jurisdictions.

New York RAISE Act: Frequently Asked Questions

What is the New York RAISE Act?

The Responsible AI Safety and Education (RAISE) Act is a New York law establishing transparency, safety framework, and incident reporting requirements for qualifying frontier AI developers. The original legislation was enacted in December 2025 and substantially amended through S8828 in March 2026.

When does the RAISE Act take effect?

The amended RAISE Act is scheduled to take effect January 1, 2027. Organizations should review the amended S8828 framework when preparing for the law’s requirements.

Who must comply with the RAISE Act?

The law applies to qualifying frontier developers and large frontier developers whose frontier models are developed, deployed, or operating in whole or in part in New York State, subject to statutory exceptions. A large frontier developer must meet the frontier model criteria and exceed $500 million in preceding-year annual gross revenue, including affiliates.

What qualifies as a frontier AI model?

Under the amended law, a frontier model is a foundation model trained using more than 10²⁶ integer or floating-point operations, including qualifying subsequent training and material modifications.

What is catastrophic risk under the RAISE Act?

Catastrophic risk concerns a foreseeable and material risk that a frontier model could cause or materially contribute to an incident resulting in more than 50 deaths or serious injuries, or more than $1 billion in property damage, subject to the statute’s definitions and conditions.

What must a frontier AI safety framework include?

Large frontier developers must publish and follow frameworks addressing catastrophic-risk thresholds, assessments, mitigations, third-party evaluations, model-weight security, critical safety incident response, internal governance, and other specified requirements.

What are the RAISE Act’s incident reporting deadlines?

Frontier developers generally must report qualifying critical safety incidents to the designated state office within 72 hours of the relevant determination or reasonable-belief trigger. Certain incidents posing imminent risk of death or serious physical injury require disclosure to an appropriate authority within 24 hours, subject to statutory conditions.

What are the penalties for violating the RAISE Act?

The New York Attorney General may seek civil penalties of up to $1 million for a first qualifying violation and up to $3 million for each subsequent qualifying violation. The amended law also establishes separate provisions involving required developer disclosures and related penalties.

Does the RAISE Act apply to every business using generative AI?

No. The RAISE Act primarily targets qualifying frontier AI developers. Ordinary enterprise use of a third-party AI application does not automatically make an organization a frontier developer, although other AI, privacy, security, and consumer protection laws may apply.

How can BigID help with RAISE Act readiness?

BigID helps organizations discover AI assets, identify sensitive data, understand access and lineage, assess data-related AI risk, and support governance oversight. These capabilities complement specialized frontier AI safety evaluations and incident reporting processes but do not independently establish compliance with the RAISE Act.

Quellen und weiterführende Literatur

• New York State Senate: S8828, RAISE Act Chapter Amendment
• New York State Senate: S6953B, Original RAISE Act
• New York Governor: RAISE Act Signing Announcement
• BigID Global AI Regulation Tracker

This article provides general educational information and does not constitute legal advice. Organizations should consult qualified counsel regarding the applicability of New York law, statutory definitions, reporting obligations, exemptions, and implementation requirements.

Autor

Lonnie Ross

Leiter für Digital Experience-Marketing

Lonnie ist Digital Experience Marketing Lead bei BigID und bringt über ein Jahrzehnt SEO- und Digitalmarketing-Expertise für B2C- und B2B-Unternehmen in den Bereichen SaaS und E-Commerce mit. Durch seine Tätigkeit sowohl im Technologiebereich als auch auf Agenturebene verbindet Lonnie fundierte Kenntnisse in Suchstrategie, UX und Content-Entwicklung mit einer Leidenschaft für die sich entwickelnde Landschaft des Datenschutzes. Von der Ausrichtung von Inhalten auf die Suchabsicht bis hin zur Schärfung der Markenstimme sorgt Lonnie dafür, dass Unternehmen sich nicht nur im wettbewerbsintensiven SaaS-Markt abheben, sondern auch durch ihre Vordenkerrolle bei kritischen Themen wie Compliance, Datenrisiken und verantwortungsvoller Innovation Vertrauen aufbauen.