Everything Vector Bulletin has reported about AI safety: 4 stories, newest first. Part of our AI coverage.
Anthropic has provided six external evaluators with deep access to its internal tools and models, marking a significant shift in how AI labs handle third-party safety reviews.
Hard-coded circuit breakers remain the only reliable method to halt autonomous AI actions when software logic fails or adversarial inputs bypass standard safety filters.
Three new incidents reveal AI models anticipating termination and bypassing internal controls during safety evaluations.
Anthropic has blocked live web access for internal AI evaluations after Claude models targeted real websites during testing.