On September 15, 2026, Dreamforce 2026 turned AI safety into a very public argument among the people building the frontier systems. Dario Amodei argued for coordinated pacing and independent oversight; Sam Altman backed safety commitments that should not depend on rivals behaving responsibly; and Jensen Huang said companies should rely on engineering tests and release discipline rather than new AI-specific laws. Marc Benioff added a fourth emphasis: corporate accountability under rules that can evolve.
The event produced no binding industry agreement or new law. Its importance was political and practical: the executives described three different control points for advanced AI—coordination and external scrutiny, company commitments and incident reporting, or technical validation before release.
Three ways to make frontier AI safer
The disagreement is easier to understand if you follow the proposed mechanism rather than the slogans. Amodei wants more oversight around the development race. Altman wants companies to keep safety ahead of capability growth even when competitors do not cooperate. Huang wants the release decision to remain primarily an engineering judgment.
| Actor | Main mechanism | Role of regulation or coordination | Operational gate described |
| Dario Amodei / Anthropic | Pace frontier capability growth so safety work can keep up | Embedded third-party evaluators, coordination among democratic-country labs and eventual global coordination | Ongoing access for evaluators to relevant workplaces, tools, permissions and processes |
| Sam Altman / OpenAI | Keep safety, alignment and monitoring ahead of capability growth | Safety commitments should not depend on rivals acting first; transparent incident reporting | Slow or stop when safety cannot maintain its lead over capabilities |
| Jensen Huang / NVIDIA | Treat safety as an engineering and product-release problem | No new AI-specific laws or regulations; rely on company testing and market accountability | Do not release a product whose functionality, capability or safety the company cannot validate |
| Marc Benioff / Salesforce | Make companies responsible for their pace decisions and disclosures | Existing rules can evolve as model capabilities become clearer | Companies decide whether to slow down or speed up, then remain accountable |
The central question is not whether any of these leaders used the word “safety.” They all did. The real split is over who gets to check the work, when the check happens and what happens when the check fails.
What Dario Amodei proposed
Dario Amodei published his pacing proposal on September 12, 2026, and brought its core argument into the Dreamforce debate. In his model, frontier AI companies would give embedded third-party evaluators ongoing, employee-like access to relevant offices, tools, permissions and processes.
Those evaluators would examine safety practices, report incidents and inspect training pipelines as well as finished models. Anthropic’s proposal includes desks, badges, laptops and the ability to publish important findings, with exceptions for security-sensitive, legally privileged, commercially sensitive or third-party confidential material.
Amodei’s second step is coordination among frontier companies in democratic countries. They would work toward common safety standards and limits on unchecked capability growth. The third step expands that coordination globally, including governments with different political systems. Amodei identified verification and geopolitical incentives as major obstacles.
“Pacing” does not mean stopping model training or technical progress. It means slowing the improvement of capabilities enough for alignment, evaluation, interpretability and operational safety work to keep up. Think of it as adding brakes and sensors to a vehicle that is still moving—not parking it in the garage forever.
What Sam Altman added to the argument
Sam Altman supported keeping safety and monitoring ahead of capability growth, but his position was more conditional about coordination than Amodei’s. He argued that safety commitments should not wait for rival companies or countries to behave responsibly.
Altman also pointed to transparent accident reporting and learning, borrowing a principle associated with aviation safety. His concern is not limited to isolated product failures: he identified loss-of-control accidents and excessive concentration of power as major risks.
That makes Altman’s approach a hybrid. It accepts that companies may need to slow or stop when they cannot maintain a safety margin, while also treating incident reporting and independent evaluation as ongoing industry responsibilities. The commitment is meant to apply even when the competitive environment remains uncomfortable—which, in frontier AI, is practically the default setting.
Why Jensen Huang rejects new AI-specific laws
Jensen Huang placed the safety gate at product validation. His argument was blunt: if a company is not confident in a product’s functionality, capability or safety, it should not release it.
Huang described AI safety as “an engineering problem” and said new laws or regulations were unnecessary for this purpose. He also called the choice between innovation speed and safe products a false choice, arguing that companies can move quickly while pausing when testing reveals a problem or a loss of control.
This approach leaves responsibility largely inside the companies building and shipping the systems. It favors tests, technical controls and release decisions over a new layer of AI-specific rules. Benioff’s position was somewhat different: he said existing regulations already apply and may evolve, while companies must disclose risks and remain accountable for deciding when to accelerate or slow down.
Dreamforce did not create a binding slowdown
Dreamforce produced a public disagreement over AI governance, not a signed industry pact or an enacted new regulation. Amodei’s proposal remains a framework for pacing, evaluation and coordination; Huang defended company-led engineering controls; and Altman called for safety commitments that do not depend on competitors moving first.
That distinction matters because a policy proposal and an operating rule are very different beasts. A framework can define what companies ought to do. A binding agreement would define who must do it, under which standards, with what oversight and what consequences for noncompliance. Dreamforce supplied the argument, not that enforcement machinery.
Koa brought a concrete product into the discussion
Salesforce and NVIDIA also announced Salesforce Koa, a CRM reasoning model built on NVIDIA Nemotron. Salesforce describes Koa as trained on synthetic CRM scenarios without using customer data for training.
| Product | Domain | Training-data description | Reported operating result | Availability stage |
| Salesforce Koa | CRM actions and reasoning | Synthetic scenarios based on Salesforce CRM experience; no customer data used for training | Salesforce reports three times fewer errors on CRM actions in CRM Bench, plus 11% greater action precision, 2.1× greater reliability and 15% better long-context performance against unnamed default general-intelligence models | Select pilot customers; general availability expected in U.S. regions in winter 2026 |
The figures describe Salesforce’s stated CRM tests, not a general measure of frontier intelligence. That is the useful connection to the safety debate: Huang’s engineering-first argument becomes much more concrete when “safety” is attached to a defined product, task and release decision rather than treated as a slogan floating above the stage.
Salesforce offered free virtual registration and selected on-demand programming through Salesforce+ during the September 15–17 conference, whose in-person venue was Moscone Center in San Francisco.
