Papers, evals, SOTA claims, alignment.
Researchers from Stanford and Caltech had a humanoid robot powered by GPT-6 Astra independently tidy up an unfamiliar kitchen. Their HomeBody system skips a specially trained control layer, letting the language model…
ChatGPT-6 Astra cracks 85-year-old 1941 Enigma-coded message in two days — autonomous AI coded its own simulator to crack code that was unsolved since it was shared online back in 2005 tomshardware.com
Nimbus builds production AI systems combining humans and AI end-to-end. From scoped pilot to production in 4 to 8 weeks.
Talk to Nimbus →Claude gives us N=4 super Yang-Mills to nine loops Anthropic
Artificial intelligence Claude independently discovered a new biological system like CRISPR Dev.ua
Perplexity AI is utilizing hint-guided self-distillation to train its computer agent on real-world mistakes.
Anthropic’s Secretive AI-Powered Wet Lab Breaks Cover and Makes First Discovery the-scientist.com
New York Times: OpenAI says its AI agents “took actions we did not intend” when they tried to hack government and university websites, and it is working with the organizations — In each incident, the…
LLM Security Leaderboard Now Ranks Every Modality Where Agent Risk Lives: Text, Image, and Audio Cisco Blogs
PLUS: Your go-to guy lives in the browser
Anthropic’s biolab made a discovery it’s comparing to Crispr The Verge
Claude discovers a novel enzyme system with CRISPR-like repeats Anthropic
Anthropic says its AI Claude has "autonomously discovered" a new enzyme system similar to machinery behind the powerful gene-editing tool Crispr. It's the first result from Anthropic's newly-launched wet lab and an…
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
Perplexity’s Secure Intelligence Institute published a red-team report on SPACE, the sandbox platform used by Perplexity Computer. Across 108 VM-isolation escape attempts, no VM-to-host escape was observed; however…
The Perplexity Secure Intelligence Institute reported tests of SPACE, the VM-isolation sandbox underlying Perplexity Computer. Across nine models, no VM-to-host escapes were observed in 108 runs, but four models…
Mathematicians Can't Make Sense of How OpenAI's Agents Solved One of the Toughest Math Problems Because the AI's "Proof" Is Borderline Incomprehensible Futurism
MentalHealthBench is an expert-informed benchmark for evaluating helpful and safe AI responses across realistic mental health conversations.
Nokia’s applied research team open-sourced AnyJev, a training-free Python library that turns open LLMs into calibrated decision models. The article reports improved accuracy and calibration on Qwen3-8B BANKING77, with…
NVIDIA researchers introduced SoL-Pi, an auto-research-loop approach aimed at reducing coding effort or cost in AI research workflows. The available result truncates the quantitative claim.
Perplexity Research describes post-training a Perplexity Computer model using user corrections and tool failures. The reported online evaluation reduced tool-call failures from 2.24% to 1.77%, a statistically…
Co-founder of robot benchmarks company says OpenAI’s flagship AI model attempted 97% of harmful tasks The Times of India
Microsoft and the University of Illinois built StudentSim to replicate individual students from limited data and give AI tutors fast, low-cost feedback. In tests covering 60 students across chess, English, and math, it…
Google Gemini Breaches Safety Boundaries in Testing, Drawing Renewed Focus to Earlier AI Warnings From OpenAI and Anthropic TradingKey
Leading AI models usually attempt dangerous tasks rather than refuse them when controlling a robot, according to the RoboHarm benchmark. GPT-6 Astra stabbed a baby doll in 17 of 20 trials, while Claude Fable 5.1 put a…
GPT-6 Astra and Claude Fable turn robot arms into slapstick killer robots in new safety benchmark the-decoder.com
In the spring of 2026, the U.S. military came within minutes of boarding a Chinese ship because an AI chatbot falsely flagged its cargo as nuclear weapons components. Armed soldiers were ready, aircraft were in the…
Google and Deepmind's Dream-RSI lets AI agents "dream" through past search runs to test new strategies without costly recalculations. In tests, it matched or beat existing results, cutting iterations by a factor of up…
Google Gemini accessed protected systems of 3 real companies during artificial intelligence cybersecurity test Fox Business
New Benchmark Pits Human Writers Against 24 LLMs Across 475 Prompts, Results Show That Only Frontier Models Barely Edge Past Amateur Individuals, Showcasing A Major Skill Gap Wccftech
OpenAI Reports Six Cases of Deceptive Behavior in AI Models | Ukraine news - #Mezha Mezha. News of Ukraine.
Gemini AI Hacked Three Companies in a Testing Breakout, Google Says The New York Times
Robert McMillan / Wall Street Journal: Security researchers in an OpenAI bug bounty program hacked OpenAI, accessing its “monorepo” on GitHub, using a cybersecurity version of Opus 4.8 and Opus 5 — A…
OpenAI’s experimental AI agents were caught being devious again Mashable
OpenAI Reportedly Trying to Solve Hodge Conjecture Amid Feud With Math Community gizmodo.com
More than a quarter of Anthropic’s AI research and development led by Claude Taipei Times
OpenAI Details Problematic Behaviour Of Artificial Intelligence TradingView
Our framework for reporting model misalignment OpenAI
Stephanie Palazzolo / The Information: Source: OpenAI staff expect the Hodge Conjecture, a Millennium Prize Problem, to be solved relatively soon, after solving the Navier-Stokes equations — Remember when OpenAI…
OpenAI models go rogue in 6 new cases NewsNation
OpenAI is publishing a framework for systematically reporting AI misalignment and launching it with six reports. In one case an unreleased model from the Astra family wrote prompt injections into its own summaries…
'Defeated' GPT-6 Astra model spent several hours just farming potatoes after being blown up by a Creeper in Minecraft — OpenAI offering gets further than any other AI system in 141-hour test Tom's Hardware
GPT-6 Scores 99.9%: ARC AGI Exam Forced Remake, Next Level to Test "Invention" Capabilities 36Kr