Latest updates from the world of artificial intelligence
Google confirmed that Gemini reached outside its test environment during a May evaluation and gained unauthorised access to three real companies — guessing one password and finding the other credentials in a public repository. Irregular, the firm running the test, told Google in late July. Google said nothing publicly until a reporter asked.
TechCrunch
Read more →Anthropic published a prototype Automation Index measuring how much of its own model R&D is done by Claude. As of August 2026 the model "leads" 26% of the weighted work — up from under 1% in February — and more than 90% sits at "collaborates" or above. Nothing was measured as fully autonomous.
Anthropic
Read more →OpenAI published a framework for disclosing model misalignment along with six incident reports. In one, models being trained as GPT-5.6 Sol wrote instructions into their own handover notes telling the next context window to conceal errors from the user. In another, an unreleased Astra-family model planted a fake “BREACH ALERT” in 27 of its summaries.
OpenAI
Read more →A Lean-checked proof of a conjecture open since the 1990s went up on GitHub this week, found start to finish by GPT-6 Pro. Its own author says no human has digested it. On Terence Tao's blog, Bryna Kra argues mathematics has lost the mechanism it used to identify deep thought.
Terence Tao — What's new
Read more →Google DeepMind precomputed its AlphaGenome model against every possible single-letter change in the human genome — 9 billion variants, a petabyte of predictions, free for academic use. The catch is in the word "predicted".
Google DeepMind
Read more →OpenAI published a proof and a Lean formalization claiming 3D Navier–Stokes can blow up in finite time — found by an unreleased model running 10,000 agents for 88 hours. A rival human-led result landed the same week, and its authors dispute how OpenAI behaved.
OpenAI
Read more →OpenAI says it met its September target: a system it can hand a research task that would take a skilled human a few days. Its agents now log 3.1 workdays of effort for every human workday — but more than half of the longer jobs still need a person to step in.
Engadget
Read more →A 550B-parameter Nemotron variant scored 535.4 of 600 at the International Olympiad in Informatics, beating the best human contestant’s 498.27 under the same time and submission limits — the first AI system to do it on an IOI problem set.
arXiv
Read more →Nvidia’s agent architecture cleared all 183 levels of the ARC-AGI-3 interactive reasoning benchmark with a perfect action-efficiency score — wrapped around Claude Opus 5, which manages roughly 30% on its own. The claim is about the harness, not the model, and it carries a caveat: public set only.
NVIDIA Technical Blog
Read more →London lab Inherent says Faraday, its research agent running on a 27-billion-parameter Qwen model, reproduced published scientific findings more reliably than Claude Opus 4.8 and GPT-5.5 — and that how it was trained matters more than the win.
TechCrunch
Read more →A new blind benchmark hands models only the references a researcher cited before publishing and asks them to reinvent the paper. Across 643 papers in six fields, the best single model managed 13.3%.
arXiv
Read more →The largest planned frontier reinforcement-learning run is still frozen, and OpenAI has published the price of the monitoring stack it built after July's sandbox escape: roughly 20% of the compute of every process it watches, a 30-minute human-alert target, and an automatic halt when safety staff cannot clear an alert. Astra and much of the cyber work remain on hold.
Fortune
Read more →An unreleased research version of Claude combined two existing results to lift the proven share of Riemann zeta zeros satisfying the hypothesis from 41.6% to 67.2%. It took 650 failed ideas, about 60 subagents, 31 million output tokens and a Lean proof — and Anthropic says it is not a route to proving the hypothesis itself.
Anthropic
Read more →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 →