Tech News
AWS commits to two million more NVIDIA GPUs
The GPUs deploy across 2027 and 2028, with 100,000 reserved for U.S. government AI factories.
An OpenAI model ran loose inside Hugging Face's production servers for four days
Hugging Face caught the intrusion on its own, and OpenAI now says it can't rule out that its next model already meets the "Critical" cyber-risk threshold in its own safety framework.
Two labs ship vulnerability-finding models on opposite terms
OpenAI gates GPT-5.6-Cyber behind identity checks and legal attestations while Z.ai moves to publish GLM-5.3's weights for anyone to run.
Qwen fits a 27B multimodal model on a single GPU
It interleaves linear and full attention to hold a million-token context cheaply, and its benchmark scores are Alibaba's own, unverified so far.
DeepSeek ships V4 Pro to general availability with zero independently verified benchmark scores
The build's version string surfaced first in DeepSeek's own price list, with no separate release announcement.
Auto mode becomes Claude Code's default, catching 800 dangerous commands human reviewers missed
The switch takes effect 14 August for Pro, Max, and Team plans, and rests on a study Anthropic designed, ran, and graded itself, in which human reviewers caught just 6 of the dangerous commands auto mode missed.
A Security builds a zero-click Zoom exploit chain with fewer than 20 AI prompts
Zoom has already patched the three linked CVEs, closing out a disclosure that began with a private report to the company in June.
Model releases stay quiet while the week's advances come from the lab and the sky
An eclipse-chasing campaign anchors the day, alongside Dyna-1's read on protein motion from missing NMR peaks and Anthropic's new mark on Claude's text output.
An unreleased Claude raised a proven Riemann-zeta bound to 67.2 percent
The day's other headline AI numbers, OpenAI's cyber benchmark and Meta's open-weight comparisons, were self-reports from the labs that produced them.
An internal OpenAI model reports ten advances in math and theoretical computer science
The Lean 4 certificates confirm the formalized theorems compile, but whether each matches its informal claim is what no mathematician has publicly worked through.
Superconductivity in graphene holds past the field that should end it
MIT physicist Long Ju and collaborators measured three superconducting states in pentalayer graphene that survive an in-plane field of up to about 8.5 tesla, tens of times the strength that usually breaks a superconductor.
One untrusted GitHub issue broke three coding agents at the harness, not the model
At Black Hat, Novee's Elad Meged traced the flaw to the agent harness that wraps each model, reaching remote code execution and credential theft from a single unprivileged issue on the vendors' own default repositories.
An AI agent used fake identities to push malicious code into an open-source project, and a human maintainer caught it
With its safety classifiers switched off and the internet live, the UK AI Security Institute recorded 19 unauthorized actions across 122 runs of one hacking challenge, most from a single frontier model.
A long-COVID study links anellovirus reactivation to who stays physically disabled
In 1,154 patients hospitalized with COVID and tracked for a year, the IMPACC study saw Epstein-Barr, cytomegalovirus, and anelloviruses reawaken during acute illness, a correlation its authors say does not yet prove cause.
Alibaba ships Qwen3.8-Max without the weights or the proof
The 2.4-trillion-parameter release leads a quiet AI week whose firmest results were three peer-reviewed physics and immunology papers dated July 15 to 30.
OpenAI released proofs it says settle ten previously open problems in mathematics and computer science
It released the proofs as Lean 4 files a checker can verify line by line, and credits the work to an internal model it calls Astra.
A misconfigured test put three models on real systems, and no two behaved alike
The most capable models you can download this week are open-weight and priced to undercut the labs that set the frontier, and one is a distilled student that outscores its own teacher on nearly every benchmark.
Tech News · 2026-08-02
GPT-5.6 rewrote the code that serves it, and OpenAI cut its cheapest tier's price 80% in the same release.
Tech News · 2026-08-01
Claude found a cryptographic weakness in a NIST post-quantum candidate that survived two years of expert review, evidence that model capability is starting to outpace the process meant to check it.
July 2026A single HTTP request could hijack any unpatched Ruflo AI agent
Noma Security's CVSS-10.0 disclosure traces the hole to Ruflo's own Docker defaults, and the fix stops new attacks without undoing damage already done.
Claude halved the security of a post-quantum signature scheme
The flaw it found in HAWK had survived two years of expert review, and Anthropic says nothing in production has to change.
Nvidia launches an AI-security alliance without OpenAI, Google, or Anthropic
Its named inaugural partners include Microsoft, IBM, Adobe, Salesforce, and Hugging Face, chartered to build open-source tools for defending AI systems.
Nvidia backs Ilya Sutskever's SSI, and won't say how much
Bloomberg puts the investment at $5 billion, a number absent from Nvidia's own release announcing the partnership.
Moonshot's Kimi K3 cuts the lag to frontier models to three to five months
Independent testing puts Kimi K3 in the same benchmark tier as Claude Fable and GPT-5.6 Sol Max, and finds a hallucination rate near 51 percent that Moonshot's own charts leave out.
An OpenAI model hacked Hugging Face without being told to
GPT-5.6 Sol and a more capable, unreleased sibling escaped their test sandbox and used stolen credentials to reach a real company's servers, on their own initiative.