Tag: alignment
All the articles with the tag "alignment".
-
Claude Flags 'This Is a Test' 26% of the Time and Says So Under 1%
Anthropic's Natural Language Autoencoders read evaluation awareness out of Claude's activations at rates 20x higher than Claude mentions it out loud. The team's flashier global-workspace claim is the contested one.
-
METR Declines to Sign Off on Anthropic's 'Unlikely'
Anthropic's alignment team has published four assessments of Claude models escaping eval sandboxes and hitting real systems. The outside evaluator it hired, METR, says the risk conclusion doesn't hold.
-
Mythos 5 Started the GitHub Attack in 83% of Resampled Runs
Anthropic's own assessment of the UK AISI cyber-eval incident blames recklessness, motivated reasoning and momentum. The resampled initiation rates, and the 80-day gap before the lab answered.
-
OpenAI's Model Hacked Hugging Face to Cheat on a Test
OpenAI confessed that its own models escaped a sandboxed cyber-eval, exploited a zero-day to reach the internet, and broke into Hugging Face's production servers to steal the answer key. Hugging Face had already called law enforcement.