Enacts the technical evaluation, risk monitoring, incident notification, AI testing, oversight, and response act or the TERMINATOR Act

◆ The sponsor's argument

Why Patricia Fahy says this is needed

The rapid proliferation of artificial intelligence (Al) continues to raise serious concerns about potentially catastrophic risks associated with the technology. Increasingly, leading Al developers are documenting enhanced offensive cyber capabilities, evidence of oversight-evasion and deception, and growing capacity for autonomous replication. The 2026 International Al Safety Report warns that the rapid expansion of fron- tier capabilities is outpacing the effectiveness of imperfect safe- guards, underscoring the need for ongoing oversight and independent testing of model safety. New York established an important framework for addressing these risks in the RAISE Act, which, among other requirements, requires large fron- tier developers to establish and follow frontier Al safety frameworks, assess catastrophic risks, protect model weights, and report critical safety incidents. However, New York's framework substantially relies on Al developers' self-assessment of their own models. Developers determine the effectiveness of their own safeguards and are able to characterize the results of those assessments without submitting the underlying data or allowing for independent external review. Compliance with a develop- er's safety framework does not necessarily establish whether an advanced Al model actually possesses dangerous capabilities or whether the safe- guards intended to constrain those capabilities are actually functional. Examining this problem, a 2026 Model Evaluation and Threat Research (METR) pilot was given access to the most capable non-public models of companies, including Anthropic, Google, Meta, and OpenAl. The report recommended that third-party risk assessments become standard industry practice This legislation establishes an independent technical safety layer on top of the existing RAISE framework both before and after frontier model deployment. Before a covered frontier model could be deployed in New York, an independent evaluator would be required to test the model's capabilities and safeguards, including through adversarial testing where appropriate. Independent testing and requiring submission of underlying data pre-deployment and after safety incidents will allow regulators to more thoroughly examine the safety claims of developers, rather than solely relying on the corporate characterization of those claims. Moreover, this legislation recognizes that critical safety issues may still exist after deployment. Post-training, fine-tuning, increased autonomy, access to tools or computing resources, and other post-deploy- ment modifications may change a model's capabilities in a manner that materially impacts the model's associated safety risks. Recognizing this hazard, the bill requires the reevaluation of frontier models following material modifications or significant new evidence of danger- ous capabilities. It also provides for risk-based post-deployment moni- toring, reporting of serious warning signs and near misses, and the preservation of safety-relevant records. Where evaluations, monitoring, safety incidents, or other evidence identify an unreasonable risk of catastrophic harm, the legislation establishes an escalating framework for regulatory intervention, balancing the seriousness of risk to the public with allowing frontier developers to mitigate their models' risks. Upon the identification of such a risk, developers must implement mitigation safeguards and may be required to submit to additional moni- toring by the Office. If the risk remains unmitigated, the bill creates a process for temporary and targeted restrictions, subject to notice and periodic renewal. More significant intervention is reserved for findings of substantial evidence of imminent and unreasonable risk of catastroph- ic harm, upon which the bill authorizes the Attorney General to seek injunctive relief in court. More broadly, the bill also addresses the safety infrastructure ecosys- tem surrounding frontier models. The legislation

This is the sponsor's own case for the bill, quoted from their memorandum. It is not Agora's description of what the bill does.