NEW YORK, November 22, 2026 — Dan Hendrycks, director of the Center for AI Safety, called for an indefinite pause on the largest AI training runs at the DealBook Summit, branding Anthropic “reckless” and demanding blunt interventions to close security flaws before further scaling.
The Shift is a Blunt Intervention
The trend underneath the current debate is a move toward more direct action. Hendrycks, who previously favored narrow technical rules, said companies have blocked such efforts for years. “So now it’s, we’re going to need a more blunt intervention. That’s how it is,” he stated. This shift means a focus on foundational safety before the next leap in capability, not incremental improvements alongside it.
What’s Changing: A Moratorium on the Biggest Models
Hendrycks’s core demand is an indefinite pause on the biggest training runs. This is not a full stop on AI development. The plan includes carve-outs for narrow models with clear public benefits, such as DeepMind’s AlphaFold for protein folding and specialized voice models. During the pause, labs could continue to sell inference services—the act of running pre-trained models for users. The focus, however, would be on hardening existing systems against known vulnerabilities.
Watch This Space: The Safety Benchmarks
The specific safety benchmarks Hendrycks demands are concrete and immediate. Before labs are allowed to scale their training runs further, he requires them to close jailbreaks and backdoors in their models. Jailbreaks are prompts that allow users to bypass an AI’s safety guidelines. Backdoors are hidden vulnerabilities in code. He also demands that companies help patch critical infrastructure and improve the overall reliability of their systems. This means fixing the flaws that allow models to be manipulated or to fail in predictable ways.
The Industry Pushback and the Political Window
The call for a pause faces significant pushback. Max Hodak, co-founder of Neuralink, argued that enforcing such a pause would require “an exhaustive surveillance state” on the internet, a regime he described as draconian. His alternative is to use antitrust law to limit what frontier labs can sell, restricting them to intelligence services and preventing them from moving into markets like drugs and cars. Hendrycks countered that the political landscape is shifting, pointing to the recent US midterms where he said voter concern about AI gave politicians room to discuss pausing development openly. He expects 2027 to be “a very interesting year for regulation.”
Sources
- Hendrycks Calls Anthropic Reckless on DealBook AGI Panel | AI Weekly
- Dan Hendrycks
- DealBook AGI panel: Dan Hendrycks wants frontier AI training paused
- Hendrycks says AI safety groups backed MisAnthropic in private—but feared signing
- Coalition challenges Anthropic safety claims ahead of IPO
Photo: TechCrunch (CC BY 4.0)





