What happened
New York Senators Andrew Gounardes and Liz Krueger Fahy introduced S10731 on October 28, 2026. The bill was referred to the Senate Rules Committee the same day. Its stated purpose is to amend provisions governing the training and use of artificial-intelligence frontier models under New York’s Responsible AI Safety and Education, or RAISE, Act.
The central proposed change is scope. S10731 would lower the training-compute threshold used to identify a frontier model from 10^26 to 10^25 integer or floating-point operations. That is a tenfold reduction in the numerical threshold. If the proposal became law, developers whose training runs meet that lower line could fall within the statute’s frontier-developer obligations.
The bill would require each large frontier developer to retain a “frontier examiner” for ongoing assessment. The examiner would produce catastrophic-risk reports by April 1, 2028, and every 90 days after that. The supplied bill material also describes reporting on the extent to which AI research and development has been automated and on the share of compute used for models deployed internally.
S10731 also revises core definitions and reportable events. Its catastrophic-risk definition retains harms exceeding 50 deaths or serious injuries, or more than $1 billion in property damage or loss, arising from a frontier model. It enumerates expert-level assistance for chemical, biological, radiological, or nuclear weapons; certain cyberattacks and crimes conducted with limited human oversight; attempts to seize government control; and evasion of developer or user control.
For critical safety incidents, the proposal includes unauthorized access to, modification of, or exfiltration of unreleased frontier-model weights. It also addresses a model’s attempts outside a controlled evaluation to circumvent, disable, or evade material technical or organizational controls that limit access to systems, resources, or data; allow suspension or modification; or monitor conduct or capabilities. Other listed cases include unauthorized impairment of external computer systems or networks, and safeguard failures that create a substantial likelihood of such an incident absent intervention, mitigation, or chance.
The proposal is not law. The bill text is a legislative draft, and its immediate recorded status is referral to committee. The supplied material says it would also authorize an office within the Department of Financial Services to order limits on training compute to avert catastrophic risk and would increase potential attorney-general penalties; those provisions would matter only if the legislation advances and is enacted.
What the archive adds
This is a new legislative event, but it belongs beside earlier reporting on Anthropic’s draft safety bill, which described a proposed judicial route for stopping a model release over specified severe harms. S10731 instead proposes state-level scope, examination, reporting, and enforcement changes to an existing New York framework.
Trend signal
The proposal supports the pattern that verification bandwidth is the scarce engineering resource. Its mechanism is regulatory rather than product-led: recurring external examination and 90-day catastrophic-risk reporting would require evidence, reviewers, and repeatable evaluation processes—not merely a written safety commitment.
Why now
The bill is framed as an amendment to the RAISE Act, which the text says was added to New York’s general business law in 2026. Its introduction follows that statutory baseline and seeks to alter definitions, coverage, assessment requirements, and enforcement tools rather than create an unrelated framework.
Connections to other events
Anthropic’s draft legislative proposal centered on whether a judge could halt a release over defined severe harms. S10731 approaches a related problem earlier in the lifecycle: which developers are covered, what safety conditions count as incidents, and what evidence must be generated on an ongoing schedule.
Why it matters
For teams building or operating very large models, the practical implication is that safety work may need to become legible to an outside examiner. A threshold change can alter who must prepare governance artifacts. Recurring reports can turn evaluations, access controls, incident response, model-weight security, and records of automated R&D into compliance inputs.
Even teams below the proposed threshold should treat this as a signal about the kind of operational evidence policymakers may seek. A policy, model card, or one-off red-team exercise is not necessarily enough to demonstrate how controls operate over time. The relevant capability is the ability to show what controls exist, what they constrain, how failures are detected, and what changed after a finding.
What happens next
- The bill receives committee action, amendment, or no further action. (3 months, 55%). Watch for: an updated action record or substitute bill.
- Debate focuses on the 10^25 threshold and examiner requirements. (6 months, 65%). Watch for: revised statutory text or committee materials.
Limits and uncertainty
This source establishes that S10731 was introduced and referred to committee; it does not establish enactment, implementation, or how regulators would interpret disputed terms. The supplied material does not identify who may qualify as a frontier examiner, what examination methodology would be required, or the operational cost of compliance. It also does not show which developers would cross the proposed threshold in practice. Requirements and penalties may change during the legislative process.
Skill to build now
Create an evidence map for a high-risk model system this week. List the technical and organizational controls that govern access, monitoring, suspension, weight handling, and incident response; assign an owner and a retrievable artifact to each. Then run one tabletop scenario in which a control is bypassed. This does not predict legal coverage, but it makes future examination and internal review more concrete.
Reader takeaway
S10731 is a proposal, not a new obligation. But its direction is clear: New York may seek to cover more frontier-model development and require repeatable, externally inspectable proof that catastrophic-risk controls and incident processes work over time.
Open question
Will lawmakers preserve the lower 10^25 compute threshold and the recurring frontier-examiner model as the bill moves through committee?