Anthropic needs to show that its models continually get better. If the model showed minimal to no improvement, it would cause significant damage to their valuation. We have no way of validating any of this, there are no independent researchers that can back any of the assertions made by Anthropic.

I don’t doubt they have found interesting security holes, the question is how they actually found them.

This System Card is just a sales whitepaper and just confirms what that “leak” from a week or so ago implied.