AI models can acquire backdoors from surprisingly few malicious documents

[CommunityNews]

Anthropic study suggests “poison” training attacks don’t scale with model size.

Read in full here:

AI models can acquire backdoors from surprisingly few malicious documents