Signs of individual dimensions in transformers carry semantic information and enable feature detection without training or rotation, opening a new path to mechanistic interpretability.
STARE uses surprisal metrics and selective advantage reweighting to maintain policy entropy stability across long training sequences while improving accuracy by 4–8%.
Claude Opus 4.7 performs complex robotics tasks without human assistance 37 times faster than human teams from a year earlier and writes code that works correctly on the first attempt in most cases.
Attackers systematically exploit legitimate AI tools and popular developer infrastructure as attack vectors while deliberately minimizing traditional security signals.
Federal government’s open-source AI model automates the retrieval of applicable law and its application to infrastructure projects to reduce approval times.
Orphaned AI agents in enterprise networks pose significant security risks because their authorization and access rights are often undocumented and not traceable.
56 percent of companies operate or plan productive AI inference in private cloud, while public cloud usage declined by 15 percentage points globally; Germany saw a more pronounced drop of 24 percentage points.
AI agents as active system participants with data access require new security approaches beyond classical governance, as their risks stem from gradual behavioral changes and Shadow AI, not from obvious violations.
Estonia’s identification number system for AI agents creates traceability of authorities and will serve as a blueprint for regulatory requirements in other jurisdictions.
RepSelect isolates forget-set-specific representations through selective gradient component collapsing and achieves 4-50x greater robustness against relearning attacks than existing methods.
