What Is Anthropic's Access Program?
Angela Jiang says Anthropic's access program lets an approved user run advanced cyber capabilities on a model such as Opus. A defensive user applies with a use case, and approval diminishes the safeguards on that user's version of the model so the user can pursue the cyber activities. Anthropic would not loosen safeguards for everyone.
The team built the program to hand specific programs out to customers. Each approval needs a per-use-case review and is scoped to a particular user. While designing the verification-style portal and the grants for specific model capabilities, they chose a more generic design because they did not know whether the same pieces would serve other jobs. The generic primitives later handled access to the model writ large, every trusted access program, and de-deployments of models.
A normal product approach would have focused only on the use cases already in view. The broader design is something Anthropic needed anyway, and the access program was the forcing function. They reach for the same solution more often than not, and it pays increasing dividends.
How Anthropic ships the #1 LLM in the World safely: Separating the Brain from the Hands ·
From the transcript
And then we have increasingly found that more often than not, by taking that strategy of product development. Most of the time, we actually find ourselves being like, we actually do use the thing. FRANCESC CAMPOY- Is there another example that's a separate one, just so I can understand? MELANIE WARRICK- For example, this access program. This access program is a really good one.
Internally, it has a bunch of granting and all these kinds of things. We developed it specifically for reasons about being able to actually give programs like this out to customers, very, very specifically. FRANCESC CAMPOY- Do you want to give a high level what the access program specifically is? Sure. So if you want dual use, let's take a model.