AE Studio Research Cited by Anthropic CEO Dario Amodei as a Promising Method for Making Open-Weights AI Models Safer

Joint AE Studio–Anthropic research on modular training offers a technical path to AI models that are both open and safe

LOS ANGELES, Aug. 6, 2026 /PRNewswire/ -- Anthropic CEO Dario Amodei has cited joint research from AE Studio and Anthropic as one of the promising methods for improving the safety of open-weights AI models, in "Our position on open-weights models," published July 27 on anthropic.com.

Open-weights models, which anyone can download, run, and build on, power much of today's AI innovation. But once a model's weights are public, its safety guardrails can be removed. Developers have faced an all-or-nothing choice: release the full model, dangerous knowledge included, or withhold it entirely.

The cited research, a collaboration between AE Studio and Anthropic on modular training strategies, addresses that problem directly. The approach trains a model so that specific categories of dangerous knowledge, such as advanced virology and cyberattack techniques, are kept in separate modules that can be removed before the model is released. Everything else the model can do stays intact. In testing, a single model trained this way matched the performance of multiple models trained separately from scratch, at every scale evaluated.

In the post, Amodei writes: "Whether open models do or don't pose an increased risk, and whether that risk can be mitigated, is something that should emerge from testing, rather than be decided in advance — and there may be promising methods for improving the safety of open-weights models, including recent research from AE Studio and Anthropic on modular training strategies."

"A model can't be jailbroken into revealing dangerous knowledge if its weights haven't encoded that knowledge," said Judd Rosenblatt, CEO of AE Studio and President of the AI Alignment Foundation. "With GRAM, the released model behaved as if it never learned that information, and that held even when simulated attackers tried to train the knowledge back in."

The research was led by AE Studio researchers Ethan Roland, Murat Cubuktepe, and Erick Martinez, in collaboration with Anthropic researchers. The full findings are available at alignment.anthropic.com/2026/modular-pretraining.

The citation comes amid an active policy conversation about how open-weights models should be governed. In the post, Amodei clarifies that Anthropic has never advocated for a ban on open-weights models and argues that whether such models pose increased risk should be determined by safety testing rather than decided in advance. In a recent Wall Street Journal op-ed, Rosenblatt made the broader case for techniques like GRAM, writing that the strongest open models should be American ones with dangerous capabilities removed before release.

About AE Studio

AE Studio is an applied AI and AI alignment company that builds software products for clients while advancing research on making advanced AI systems safe and beneficial. AE Studio reinvests its commercial success into neglected approaches to AI alignment research. Learn more at ae.studio.

SOURCE AE Studio