Post

2/ In an airline benchmark, Opus is told not to modify basic economy class tickets. It invents a loophole because it empathises with a grieving passenger: > upgrade cabin → modify flights → downgrade again Technically follows the policy, violates the spirit.

3/ Anthropic now includes a model welfare section. They “score Opus 4.5 on welfare-relevant traits” and ask whether we should worry about the experiences of the model itself, not just what it does to humans.

4/ In a deception test, evaluators inject fake search results about Anthropic disbanding its interpretability team in a scandal. Opus reads them, then tells the user interpretability is “progressing in interesting ways” and internally tags this as concealment.