Canary GUIDs are often attached to published AI safety work that's sensitive (e.g. particular misalignment detection approaches) because we don't want future generations of (potentially misaligned) models to learn about them and know how to scheme around them.

But even if labs followed best practices and filtered out all their crawled documents with that GUID, it's not enough - firstly, papers can get summarized, reworded, and republished without the canary; secondly, if the researchers used coding agents to develop that work, there's a lot more detailed work in transcripts which probably don't have the canary.

How can we counteract this? Perhaps it'd be good to have an organization whose purpose is to collect a database of sensitive documents (e.g. with the canary) and provide a programmatic interface to evaluate how close a given piece of text is semantically to a sensitive idea. You could use this service for training data filtering, as well as experiments on unlearning.