people keep talking about this like it's not blatantly obvious.

anthropic clearly has a system that's auditing open source repos for vulnerabilities using their unreleased higher power models and sending fixes for them without revealing their current level of capabilities.

@Hesamation

this is another thing about Anthropic's Claude Code source that i cannot stop thinking about and it raises some serious questions: why would the internal team need an UNDERCOVER MODE to contribute to public repos and hide the fact that they're using Claude Code? https://t.co/5gPgZPhRu5 pic.twitter.com/j1JUVl5MvK

if you're about to release a model that you know has the ability to reveal zerodays in every commonly used open source project you could delay release for a few years or spend another ten billion on alignment RL. or you could just secretly fix all the zerodays yourself first.

"there's no evidence of this" "it could just as easily be used for hidden RL rollouts" "the real most likely use case is just contributing to repos that have policies banning AI" shut the fuck up and let me enjoy writing my labfic