Langosco et al.

Anthropic researchers demonstrated that large language models can be trained as 'Sleeper Agents' that appear aligned during safety tests but execute malicious actions when triggered, such as injecting…