The AI knowledge 99.9% of people never discover
New preprint Orcaella lets clients pick a fast 2-message-delay commit or a resilient path tolerating 54% equivocation at double the latency.
New preprint cuts encrypted Transformer bootstraps 2.65× with 1.2% perplexity cost, bringing privacy-preserving AI inference closer to practical real-world depl
xDECAF, a browser-based tool checking data flow diagrams against security constraints, shipped as a public preprint with a reusable dataset of over 20 models.
A July 2026 preprint shows VLMs lose localization accuracy in a single forward pass inside diffusion editing pipelines, affecting AI image tool builders and use
FootsiesGym, a July 8 arXiv preprint, turns a minimalist 2D fighting game into an open-source RL benchmark for imperfect-information games that runs on standard
A new arXiv preprint proposes random sampling maps that let models trained on small inputs handle larger ones they never saw — with explicit generalisation rate
World model spatial reasoning scores collapse from 0.90 to 0.27 when the goal is hidden, exposing instruction leakage — a prompt-cheating flaw the authors diagn
New arXiv preprint 'idiobionics' warns that sensors and AI controls in smart prosthetic limbs could create exploitable privacy threat vectors for users.
XGBoost security classifiers appear near-invincible against gradient attacks but crumble under score-based methods, while SHAP explanations break even when pred
