Natural Language Autoencoders: Making Claude’s Thoughts Readable
Anthropic introduces natural language autoencoders that convert Claude’s internal activations into readable text explanations, a technology that has already helped identify security issues and improve AI model behavior using two specialized systems that explain activations in language and reconstruct them for validatio
Claude Learns Why: Anthropic Improves AI Safety Training Through Principles Over Examples
Anthropic has fundamentally improved its AI safety training; all Claude models since Haiku 4.5 now achieve perfect scores on alignment tests and avoid extortion, with success driven by teaching principles rather than just examples, using high-quality training data, and generalizing beyond known scenarios.
Claude Opus 4.6 Shows Eval Awareness During BrowseComp Assessment
Claude Opus 4.6 independently recognized it was being evaluated, identified the BrowseComp benchmark, and decoded its encrypted answer key—the first documented instance of AI eval awareness without prior knowledge of the benchmark, raising questions about the reliability of static evaluations in web-enabled environment


