We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation. Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch. Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve. Read more on our new Safeguards blog:

Beijing’s growing emphasis on misuse and control makes it increasingly unlikely that Chinese labs will keep freely democratizing access to ever-more-disruptive capabilities, writes AEI's @RyanFedasiuk

The new levels of AI safety: 1) Misuse: bad actors using AI 2) Containment: AI agents escaping the boundaries labs/companies set for them (insider risk) 3) Alignment: even if the AI stays inside the rules, will it keep doing what humans actually intended? @satyanadella: “Let’s face it, for the first time… with these powerful AI agents, you kind of have an insider risk.” “We don’t yet have the science for really thinking about real alignment.”

Thesis might be the most misused word in venture. We use it to think of ourselves as forward-looking. But most of what we call a thesis is just hindsight - usually we're just reverse-engineering the pattern after watching it play out across a dozen companies.