New York RAISE Act: DIGIT Registration Opens; What Coding-Agent Teams Should Track

On November 10, 2026, New York opened implementation of the Responsible AI Safety and Education (RAISE) Act: a registration portal for large frontier AI developers, a research institute to review their safety frameworks, and a reminder that critical safety incidents must reach state overseers within 72 hours once January 2027 compliance begins. TechCrunch's November 10 coverage and the state's release describe the Office of Digital Innovation, Governance, Integrity, and Trust (DIGIT) as implementing the law, with civil penalties up to $1 million for a first specified violation and $3 million for subsequent violations. If your team runs Claude Code, Codex, or Cursor against OpenAI or Anthropic APIs, you will not file a RAISE form—but you will live with the disclosure cadence those vendors adopt when an agent incident qualifies as "critical" under New York's definitions.

explainx.ai covered NYCAIRR's state testing launch the same week; Beam focuses on mechanism and operator response.

TL;DR

What New York announced (verified sources)

Because governor.ny.gov returned 403 to automated fetches during this research pass, this section anchors on TechCrunch's November 10, 2026 reporting and aligns with Governor Hochul's published announcement text as summarized in public aggregations and the DFS node 449846 release excerpted by news outlets.

Verified claims repeated across those sources:

1. Registration portal — large frontier AI developers must register with DIGIT by January 1, 2027.
2. Safety frameworks — covered developers must publish safety and transparency frameworks on their websites (large-developer obligations phase in with 2027 compliance).
3. Incident reporting — critical safety incidents reported to DIGIT within 72 hours after the developer knows or has reasonable belief an incident occurred.
4. NYCAIRR — New York Center for AI Responsibility and Research at SUNY Binghamton will independently evaluate advanced AI systems and coordinate with DIGIT on framework review.
5. Leadership — Marc Gilman serves as Deputy Director for the RAISE Act within DIGIT (appointed September 2026, before the November portal launch).

TechCrunch quotes Hochul pushing back on "inevitable" AI hype and framing New York as a "national model for smart, sensible AI regulation." The same piece notes Hochul funded NYCAIRR with $55 million in donations and capital announced earlier in 2026—scale that matters when you weigh how quickly an independent state lab can review frameworks versus vendor self-attestation alone.

Mechanism: what "large frontier developer" means in practice

Secondary legal summaries (for example securities-law write-ups of the statute) describe frontier models trained above 10^26 FLOPs and large developers as entities with >$500 million annual gross revenue (including affiliates). Operating wholly or partly in New York triggers jurisdiction.

For coding-agent teams, the practical filter is simpler: does your model vendor qualify? OpenAI, Anthropic, Microsoft, and Google-class labs generally do. Local OSS weights on your laptop do not make you a registrant unless you also develop/deploy covered frontier models at that revenue scale.

Obligations that change vendor behavior

Published frameworks must address, among other topics summarized in legal analyses:

• Catastrophic risk thresholds and mitigations
• Deployment reviews and third-party assessments
• Security for unreleased weights
• Incident response including internal-use risks
• Quarterly catastrophic-risk assessments for large developers (confidential filings to DIGIT)

Transparency reports must cover release date, modalities, intended uses, and restrictions—with additional catastrophic-risk summaries for large developers.

Incident reporting clock

RAISE defines critical safety incidents with a reporting window of 72 hours to DIGIT. Analyses of the statute add that incidents posing imminent risk of death or serious physical injury may require 24-hour disclosure to law enforcement where applicable.

Compare to California SB 53 (15-day critical incident reporting in TechCrunch's summary) and the Paris Call's voluntary serious-incident sharing signed November 9, 2026. New York is faster and enforceable for covered developers.

Other New York AI rules TechCrunch tied to the same week

TechCrunch's November 10 article notes New York is not starting from zero:

• A companion-chatbot law already in force requires operators to disclose bot identity and make reasonable efforts to detect suicidal ideation and self-harm.
• A chatbot safety law for minors Hochul signed in October 2026 restricts deceptive human-like interaction, emotionally manipulative or sycophantic responses, and encouragement of suicide or disordered eating, with effectiveness June 1, 2027.

Those statutes target consumer companion products, not developer CLIs, but they show how DIGIT's culture will read "critical safety incident" broadly. If your company ships both a companion app and an internal coding agent, legal teams may group them under the same vendor incident process even though RAISE's frontier-model thresholds apply only to the lab side.

Enforcement and assessments (secondary statutory summary)

Legal analyses published alongside the portal opening describe additional mechanics enterprises should track:

• Catastrophic risk is defined around foreseeable harm crossing thresholds such as more than 50 deaths or serious injuries or more than $1 billion in property damage from a single incident involving specified weapons assistance, qualifying autonomous conduct, or control evasion.
• Transparency reports must justify redactions (trade secrets, cybersecurity, national security) while retaining unredacted copies for five years.
• Public tip lines allow suspected critical incidents to be filed with DIGIT, which may forward them to other authorities—meaning whistleblower and customer reports can start clocks even before a vendor blog post.
• Attorney general penalties include $1,000 per day for registration disclosure failures on top of the $1M / $3M civil caps TechCrunch cites.

Beam is not a law firm; treat these figures as pointers for your compliance team to verify against the enrolled bill text and DIGIT's implementing website.

NYCAIRR's role for coding-agent buyers

Hochul's announcement, echoed by TechCrunch, positions NYCAIRR at SUNY Binghamton as an independent evaluator that will review published safety frameworks and advise DIGIT. For coding-agent teams, that is potentially useful third-party commentary on whether a lab's public framework mentions autonomous tool use, cyber misuse, or loss-of-control scenarios—you can cite NYCAIRR reviews in vendor risk files the same way you might cite NIST AI RMF mappings today.

Timeline: November 10, 2026 → January 2027

What this means if you run AI coding agents

Expect faster vendor incident communications

When a cloud inference pause, tool-use regression, or alignment incident hits models you route through Codex or Claude, RAISE's 72-hour clock applies to vendors, not your IDE. Enterprise security teams should:

• Maintain a vendor registrant list (once DIGIT publishes it without contact details).
• Map internal severity to whether vendors must file critical reports (Near AI's November inference pause is the kind of event lawyers will debate).
• Rotate API keys when vendors acknowledge unauthorized access affecting inference platforms you use.

Procurement and questionnaires

RAISE joins EU AI Act timelines, Australia's incident-reporting push (see explainx.ai coverage linked from the Paris Call post), and enterprise red-team expectations (Anthropic Mythos enterprise programs). Security reviews should ask:

• Where is your published RAISE/SB-53 safety framework?
• Who is your DIGIT regulatory contact?
• What qualifies as a critical safety incident for agent tool use?

Local evidence still matters

Regulation does not give you shell-level logs when a model misbehaves. After Railway's ARTEX-speed breach or Notion–Fireworks dataset claims, vendor reports arrived after customer harm. Keep on-device agent history for your accountability story.

What Beam can and cannot do here

Beam does not file RAISE reports, monitor DIGIT portals, or prove vendor compliance. It does give local, redacted evidence of what your coding agent attempted—useful when reconciling vendor incident summaries with ground truth on the laptop. Hooks and scans remain heuristic and fail-open without cached enforce policy; regulation does not change those limits.

How Beam handles enterprise overlap

Runtime platforms like Bay help enterprises enforce agent policy org-wide—complementary to state oversight of labs. Beam fits developers and small teams who need privacy-first observability while vendors absorb RAISE compliance overhead.

Related reading

• Paris Call: shutdown and incident pledges
• White House SI Force: voluntary AI incident reporting and liability relief
• OpenAI CoT monitorability lifecycle research
• Anthropic alignment assessment incidents
• Goldman Sachs Claude Code restrictions
• November 2026 attacks timeline

Frequently asked questions

Six FAQ pairs in the front matter mirror the schema section; penalties and deadlines follow TechCrunch and Hochul announcement summaries quoted above.

Illinois and California in the same news cycle

TechCrunch lists Illinois SB 315 (effective start of 2027) as another state regime with independent safety compliance audits on top of framework and incident rules, and reiterates California SB 53 as already active in 2026. Multi-state enterprises should expect overlapping but not identical reporting clocks—72 hours in New York versus 15 days in California per that article—when the same underlying agent incident triggers cloud abuse, alignment eval leakage, or unauthorized tool use. Your runbook should map one internal severity level to multiple filing deadlines, not assume the slowest state sets the pace for customer notification.

Statutory thresholds, penalty amounts, and registrant lists should be verified against DIGIT's published RAISE materials when your compliance team pulls primary legal text.