Technology & AI
DROPJ trains safe AI agents from human justifications
New arXiv paper pairs world models with human preferences and justifications to train safe AI agents without risky trial-and-error deployment.
Technology & AI
Daily AI and technology coverage — models, chips, agents and research, and what each development actually means for you and your business.
552 stories
Technology & AI
Google Vids adds text-to-video editing and selfie avatars
Google Vids now lets subscribers generate and edit video from text prompts and create a talking avatar from a selfie, but access is limited to paid tiers
Technology & AI
Adversarial prompting framework finds encoded attacks bypass AI safety
New arXiv preprint from Microsoft and Walmart researchers finds encoded prompts bypass AI safety filters most, raising the stakes for enterprise AI
Technology & AI
Agent-ready websites nearly double AI shopping agent success
A new arXiv framework lifts AI browser-agent task completion from 49% to 89% by restructuring pages for machine reading, hitting every e-commerce site
Technology & AI
OriginBlame cuts AI training over-deletion from 101x to 1.3x
A new arXiv system called ob tracks author identity through AI data pipelines, letting trainers build precise forget sets when contributors demand removal
Technology & AI
Editing AI reasoning chains cuts tokens 40%, correction 25%
arXiv preprint proposes directly editing chain-of-thought steps in LLMs, claiming 25% better error correction and 40% token savings on STEM tasks.
Technology & AI
AI upskilling framework: 3 learners pass NVIDIA Agentic AI exam
An arXiv preprint details a five-stage AI upskilling framework. Three learners passed NVIDIA's Agentic AI exam, but evidence is thin and unpeer-reviewed.
Technology & AI
NVIDIA Cosmos 3 Edge puts 4B-parameter AI on Japanese robots
NVIDIA's new on-device model lets Japanese robotics giants like FANUC and Yaskawa run vision and reasoning locally on edge hardware, adapting to new tasks
Technology & AI
AI watermarking paper maps exact token cost of tracing users
arXiv preprint reveals exact token cost of attributing AI text to a user, plus a window where generated text is provably machine-made but unattributable.
Technology & AI
Moving target defence drops attack success as low as 4%
A new arXiv framework shows aggressive network path randomisation drops attack success to 4-20% and lifts throughput 30.9% for enterprise defenders.
Technology & AI
PVDetector catches prompt injection attacks with under 1% miss rate
A new arXiv preprint proposes a training-free detector that reads LLM hidden states to catch prompt injection attacks on purpose-specific agents, reporting
Technology & AI
Claude Sonnet beats GPT-4.1 on cost despite higher token price
IBM Research found Claude Sonnet cost half as much as GPT-4.1 across 417 agent tasks, because caching matters more than sticker price for AI routing.