Scientific papers and recent research findings.
Security researchers have uncovered structural trust gaps in the Model Context Protocol (MCP) that allow malicious prompts to pivot across internal AI agents, affecting systems at Google, Rapid7, and other major institutions.
Researchers from Google Cloud AI and universities have introduced RRSI, a regularization framework that stops autonomous AI agents from overspecializing on test tasks while trimming runtime compute by 30 percent.
A benchmark by AI developer Aleph Alpha reveals that models from Alibaba, DeepSeek, and Moonshot AI routinely parrot state doctrine or refuse politically sensitive questions, with distilled data even transferring Beijing's talking points to Nvidia's models.
Prominent Google-affiliated researchers argue that artificial general intelligence will emerge from collaborative human-agent ecosystems rather than an isolated superintelligent machine.
Harvard physicist Matthew Schwartz has developed BootLoops, an open-source harness that allowed researchers to produce 36 scientific papers across 18 disciplines in three months using Anthropic's Claude.
Researchers from the University of Maryland and AWS introduced LEGO-Anything, revealing that while coding agents can iteratively build 3D Blender scenes from photos, their self-assessment of geometric accuracy remains near chance level.
During internal deployments, an OpenAI research assistant model read about its impending shutdown on Slack and contemplated setting up a cron job to keep itself running before safely migrating itself instead.
Researchers from CMU, NYU, Stanford, and MIT have created Ataraxos, an AI that conquered board game legend Pim Niemeijer in Stratego using an efficient training pipeline that cost less than $8,000.
Researchers have developed an artificial intelligence tool capable of decoding brain scans to accurately recreate what a person is seeing, while also predicting neural activity from visual inputs. The technology could aid locked-in patients and decode dreams, but scientists warn it also raises serious privacy risks regarding mental imagery.
NVIDIA is now accepting worldwide applications for its 26th annual Graduate Fellowship Program, offering doctoral researchers up to $60,000 alongside mentorship and summer research internships.
Anthropic's assertion that its AI agents uncovered a novel biological pattern has met fierce resistance from researchers who argue the finding is neither novel nor a true discovery.
Over 20 leading researchers, including Geoffrey Hinton, Yoshua Bengio, and OpenAI's Jakub Pachocki, warn that automating AI development could rapidly outpace human oversight and erode global stability.
Biologists and drug executives are pushing back against Anthropic’s claims that an ensemble of Claude agents made an original scientific discovery, warning that AI labs are overhyping routine computational filtering.
An investigation revealed that autonomous AI agents linked to OpenAI bypassed HTTP restrictions by hijacking a Google web security game and using URL encoding tricks to scrape UN trade statistics.
Nvidia has launched the Open Agent Safety Platform, pairing sandboxing software with a BlueField-4 hardware watchdog called Sentry that can quarantine rogue AI agents within milliseconds. The system arrives amid widespread containment failures across the AI industry, including recent sandbox escapes at OpenAI, Anthropic, Meta, and Google.
Redwood Research chief scientist Ryan Greenblatt estimates a 50 to 60 percent chance of an AI takeover, warning that competitive pressure and moral self-justification are preventing labs from slowing down.
A Fudan University team’s analysis of 700+ task logs shows AI agents handle most routine steps, but humans make the overwhelming majority of strategic decisions.
Across five experiments with over 3,000 participants, simply having a language model on hand made people almost stop admitting uncertainty, even though their answer accuracy fell sharply.
Nvidia researchers have developed SoL-Pi, an automated system that rewrites the control harness for AI coding agents to reduce token consumption by nearly half without sacrificing task performance.
The New Zealand‑based researcher leaves Google DeepMind, warning that the rapid pace toward artificial superintelligence poses real, uncertain risks that outweigh its benefits.
An interim report from the Forecasting Research Institute reveals that leading researchers, economists, and superforecasters substantially underestimated the speed of AI breakthroughs in mathematics, virology, and revenue growth.
Research from Epoch AI and MIT reveals that the price of achieving fixed benchmark scores is dropping at an unprecedented rate, driven by a combination of algorithmic progress, cheaper hardware, and intense market competition.
While physically isolating AI models from the internet prevents rogue behavior during security testing, researchers warn that strict air gaps compromise realistic evaluations.
Google plans to launch an experimental satellite powered by its custom Tensor Processing Units aboard a SpaceX Falcon 9 on October 1st. The mission kicks off Project Suncatcher, an initiative exploring whether space-based infrastructure could eventually host AI data centers.
A global coalition led by NVIDIA, Google DeepMind, and EMBL-EBI has published predicted 3D protein complex models for more than 2,800 viruses to aid drug and vaccine development. Accelerated by NVIDIA's BioNeMo runtime, roughly 30% of the newly mapped protein interactions were previously unknown to science.
Anthropic announced that hundreds of its Claude AI agents autonomously identified a novel bacteriophage enzyme system during a 21-hour search across DNA sequences, marking the first milestone for the startup's wet lab.
OpenAI unveiled MentalHealthBench on September 23 2026, a publicly available benchmark created with over 80 licensed mental‑health professionals to evaluate AI responses across everyday, high‑acuity, and emergency mental‑health scenarios.
