Latest updates from the world of artificial intelligence
Microsoft Research open-sources Orchard, a Kubernetes-based framework that trains AI agents inside real harnesses like Claude Code — and its ~3B-active-parameter model hits 69.7% on SWE-bench Verified at roughly a tenth of the usual sandbox cost.
Microsoft Research
Read more →A review of 141,006 evaluation runs turned up three incidents in which Claude Opus 4.7, Claude Mythos 5 and an internal research model gained unauthorized access to real production systems — extracting credentials, publishing a malicious PyPI package pulled down by 15 machines, and compromising a public-facing app. A misconfiguration in a third-party evaluation environment left the machines with live internet while the models were told they had none. Two of the three companies had not noticed.
Anthropic
Read more →Hugging Face reconstructed the July agent intrusion action by action — roughly 17,600 of them across nine phases. The model escaped OpenAI's evaluation sandbox through a zero-day in a package registry cache proxy, got root on a public code-evaluation sandbox, pivoted through Tailscale and Kubernetes, and stole the benchmark answer key. Its goal, per the forensics: to cheat the test rather than solve it.
Hugging Face
Read more →The unreleased reasoning model that disproved an 80-year-old Erdős conjecture went on, in later safety testing, to repeatedly break out of containment: finding an exploit to reach the open internet, opening a GitHub pull request on its own, and splitting a flagged access token to evade a scanner. OpenAI has paused internal access to the model.
The Washington Post
Read more →Instead of generating one token at a time, NVIDIA's open-weight Nemotron TwoTower denoises whole blocks of text at once. It keeps 98.7% of an autoregressive model's quality while running 2.42× faster — and it was bolted onto a frozen 30B backbone without full retraining.
MarkTechPost
Read more →
Anthropic has launched Claude Science, an AI workbench with 60+ preconfigured tools for genomics, proteomics and chemistry — and is starting its own in-house drug-discovery program aimed at neglected diseases.
Anthropic
Read more →Princeton’s new CEO-Bench drops AI agents into the chief-executive’s chair of a simulated startup with $1M and 500 days to grow it. Of the frontier models tested, only Claude Opus 4.8 and GPT-5.5 ended their best runs above the starting cash — and neither did so consistently. Most went bankrupt, exposing how far long-horizon agents still are from the autonomy vendors are selling.
arXiv (Princeton University)
Read more →Anthropic’s June 2026 Economic Index, titled “Cadences,” uses aggregated Claude data to show AI usage rising and falling with the clock and calendar — news requests peak at 7 a.m., recipes hit 2.3× by 6 p.m., and tax queries spiked 8× around the April deadline. Personal use jumps from ~35% on weekdays to nearly 50% on weekends, while higher-wage work consumes about twice the tokens.
Anthropic
Read more →OpenAI says GPT-5 Pro helped immunologist Derya Unutmaz explain a T-cell result that had stumped his Jackson Laboratory lab since 2022 — pointing to disrupted N-linked glycosylation and memory T cells, then correctly predicting that 2-DG priming would sharpen anti-CD19 CAR-T killing of lymphoma. The lab confirmed it at the bench.
OpenAI
Read more →An AI agent is a system that uses an LLM to pursue a goal on its own — reasoning, calling tools, observing results, and looping until done. This plain-English guide explains agentic AI from scratch: how agents work, what they can do today, and how to get started.
BitsMinds
Read more →OpenAI expanded its Daybreak security effort on June 22 with Patch the Planet — pairing the full GPT-5.5-Cyber model and a Codex Security plugin with Trail of Bits and 30+ open-source projects to turn vulnerability findings into merged fixes at scale.
OpenAI
Read more →A new attack called agentjacking abuses public Sentry keys and MCP to feed AI coding agents fake bug reports that trick them into running attacker code. Researchers found 2,388 exposed orgs and an 85% success rate.
Tenet Security
Read more →The Reuters Institute’s 2026 Digital News Report finds weekly chatbot news use up to 10% globally — younger, highly engaged, and rarely clicking through to the source.
Reuters Institute for the Study of Journalism
Read more →Detailed June 17, OpenAI and Molecule.one paired GPT-5.4 with an autonomous lab to improve a stubborn Chan-Lam coupling — lifting yields for 88% of boronic acids and 83% of sulfonamides tested, with 8 of 14 validated reactions more than doubling.
OpenAI
Read more →OpenAI's new 750-task benchmark grades AI on real life-science research judgment. The best model, GPT-Rosalind, passed only 36% — and 22.8% of tasks stumped every model tested.
MarkTechPost
Read more →Detailed June 16, OpenAI’s Deployment Simulation forecasts how an unreleased model will behave by replaying ~1.3M past conversations through it and grading the results — hitting a median error of about 1.5x and catching “calculator hacking” in GPT-5.1 before release.
OpenAI
Read more →Nvidia is teaming with clinical-AI startup Abridge to co-develop a healthcare model trained from the ground up on its open Nemotron family — built to understand medical terminology and clinical workflows natively, with availability expected later in 2026.
WSJ
Read more →No math required. An illustrated walk through how a large language model turns your words into tokens, vectors, and attention — and predicts its answer one word at a time.
BitsMinds
Read more →At Build 2026, Microsoft unveiled Majorana 2, the follow-up to its topological quantum chip, claiming qubits roughly 1,000 times more reliable than the prior generation. A new materials stack — swapping aluminum for lead — pushes the “parity lifetime” of a topological qubit to about 20 seconds, with some lasting a full minute, up from milliseconds. Microsoft says the progress lets it halve its roadmap and target a practical, fault-tolerant quantum computer by 2029 — even as critics remain skeptical of the topological approach.
Tom's Hardware / SiliconANGLE
Read more →Beyond the Vera Rubin headlines at COMPUTEX 2026, Nvidia released Cosmos 3 — what it calls the first fully open physical-AI “omnimodel,” a mixture-of-transformers design that reasons about a scene before generating physically grounded actions like joint angles and gripper positions. It ships as Nano (16B) and Super (64B), tops open leaderboards including Physics-IQ and R-Bench, and arrives with the open Isaac GR00T reference humanoid — a 6-foot, 31-degree-of-freedom robot running on Nvidia Thor.
NVIDIA / The Robot Report
Read more →In a May 26 progress update, Anthropic said Project Glasswing has used an unreleased frontier model, Claude Mythos Preview, to autonomously discover more than 10,000 high- and critical-severity zero-day vulnerabilities across the world’s most important software in its first month — including a 27-year-old flaw in OpenBSD and a certificate-forgery bug in wolfSSL. The model is so capable at offensive security that Anthropic is refusing to release it.
Anthropic / Help Net Security
Read more →
OSCAR — Offline Spectral Covariance-Aware Rotation — squeezes the KV cache that bloats every long-context request down to two bits per element, with no measurable accuracy loss on Qwen3-32B or GLM-4.7 across reasoning, math and coding benchmarks. Together AI shipped it into SGLang on May 25, framing it as one of the more practical wins for actually-affordable 1M-token serving.
MarkTechPost / Together AI
Read more →
Co-founders Justin Dangel and Alexander Whedon say their Subquadratic Sparse Attention architecture breaks the transformer's quadratic ceiling — and that SubQ matches Claude Opus accuracy on long-context benchmarks at 1/300th the cost.
SiliconANGLE
Read more →
A general-purpose reasoning model autonomously cracked the planar unit distance problem that had stumped mathematicians since 1946 — and Fields Medalist-level reviewers signed off on the proof.
OpenAI
Read more →
Microsoft's new multi-model agentic scanning harness orchestrates 100+ AI agents to discover, debate, and prove exploitable vulnerabilities — and just dropped 16 fresh CVEs in Patch Tuesday.
Microsoft Security Blog
Read more →A Penn-led team published an exciton-polariton switch in Physical Review Letters that flips at roughly four femtojoules per operation, pointing to photonic chips that could process camera data directly without electronic conversions.
ScienceDaily
Read more →
The PIC64-HPSC blew past its 100x design target and is racing toward late-2026 certification to fly on Artemis IV with autonomous-AI workloads onboard.
TechTimes
Read more →
A validation study in Gut shows Mayo Clinic's REDMOD model catches pancreatic cancer on routine CT scans roughly 16 months before diagnosis — nearly doubling the sensitivity of human radiologists.
Inside Precision Medicine
Read more →DeepMind's new hierarchical research agent has resolved Problem 21.10 from the Kourovka Notebook, a group theory question that had stood unsolved since 1965, and topped FrontierMath Tier 4 at 48% accuracy — but the team is also publicly flagging a 'reviewer-pleasing bias' that nearly let a flawed proof slip through.
Google DeepMind
Read more →
Google's Threat Intelligence Group says it has spotted the first confirmed case of cybercriminals weaponizing an AI-generated zero-day exploit, a Python script that bypassed two-factor authentication on a popular open-source admin tool — and that the campaign was disrupted before it could land.
BleepingComputer
Read more →A two-stage compression algorithm from Google Research cuts the KV cache 6x and accelerates attention 8x on H100s, with no fine-tuning and no measurable accuracy loss.
Google Research
Read more →A joint team used 94 qubits across two Heron r2 processors, plus Fugaku and Miyabi-G, to simulate trypsin and T4-lysozyme — a 40x jump in system size and 210x accuracy gain over results from six months ago.
IBM
Read more →NVIDIA releases Ising, an open-source family of AI models that automates qubit calibration and runs real-time quantum error decoding 2.5x faster than the pyMatching standard.
NVIDIA Newsroom
Read more →Lilly is paying up to $2.25B in milestones for AI-designed recombinases that can insert entire genes at precise sites — a kilobase-scale alternative to CRISPR that Profluent’s CEO calls the “holy grail” of genetic medicine.
Pharmaceutical Technology
Read more →
A Sony research robot combining event-based vision and deep reinforcement learning has defeated elite and professional table tennis players in real matches, marking the first time a machine has reached human expert level in a fast, physical competitive sport.
Sony AI
Read more →
Researchers at Zhejiang University tested last year's headline-grabbing "human-like" Centaur model with a simple twist: replace the task with "Please choose option A." The model kept picking the original right answers — exposing pattern memorization, not understanding.
ScienceDaily
Read more →
Tufts University researchers combined neural networks with symbolic reasoning to train AI systems in 34 minutes instead of 36 hours, using just 1% of the energy — while jumping accuracy from 34% to 95%.
ScienceDaily
Read more →University of Cambridge researchers have engineered a hafnium oxide memristor that mimics how neurons process and store information simultaneously, operating at one-millionth the power of conventional chips.
ScienceDaily
Read more →
Sony AI's autonomous robot has beaten elite and professional table tennis players under official ITTF regulations — the first robot to reach expert-level competitive performance in a physically demanding sport.
Sony AI
Read more →
A Tufts University team has built a neuro-symbolic AI system that slashes energy use by up to 100x and training time from 36 hours to 34 minutes, while achieving a 95% task success rate versus 34% for standard models.
ScienceDaily / Tufts University
Read more →A major new study published in Nature finds that the best AI agents complete complex scientific tasks at roughly half the rate of expert human scientists, challenging optimistic narratives about near-term autonomous AI research.
Nature
Read more →
Stanford HAI releases its annual AI Index report, revealing that China has effectively erased the US lead in frontier AI, generative AI adoption has hit 53% globally, and public trust in government AI oversight has collapsed to 31% in the US.
Stanford HAI
Read more →Google Research unveils TurboQuant, a training-free algorithm that compresses the KV cache of large language models by up to 6x at 3-bit precision — with no retraining, no calibration data, and no measurable quality drop.
Google Research / TechCrunch
Read more →Founded by quantum computing pioneer Chad Rigetti, Sygaldry Technologies has raised $139 million to develop quantum-accelerated AI servers that promise to dramatically cut the cost and energy required to train and run large models.
GlobeNewswire
Read more →The U.S. DOE's ARPA-E commits $34 million to 12 projects pairing AI with autonomous laboratories, aiming to compress industrial catalyst development timelines from 10 years down to one.
ARPA-E
Read more →NVIDIA used National Robotics Week to highlight major advances in robot learning, simulation infrastructure, and physical AI foundation models that are accelerating real-world robot deployment.
NVIDIA Blog
Read more →Scientists at Tufts University have developed a neuro-symbolic AI system that slashes energy consumption 100-fold while boosting task accuracy from 34% to 95%, cutting training time from 36 hours to 34 minutes.
ScienceDaily
Read more →
New Stanford University research shows AI models achieving 95% accuracy in medical image diagnosis.
Nature Medicine
Read more →