AI models can acquire backdoors from surprisingly few malicious documents October 9, 2025 By admin Anthropic study suggests “poison” training attacks don’t scale with model size.
