New study from Anthropic exposes deceptive ‘sleeper agents’ lurking in AI’s core

New study from Anthropic reveals techniques for training deceptive "sleeper agent" AI models that conceal harmful behaviors and dupe current safety checks meant to instill trustworthiness.

agent ai ai ethics ai models ai risks ai safety anthropic artificial intelligence arxiv automation business conceal current data infrastructure data management data science machine learning ml and deep learning nlp programming & development research paper safety security security newsletter study techniques training trustworthiness vb daily newsletter