demis out as ceo and jeff dean gone after 27 years, same day ... that is a lot of institutional memory leaving one lab at once. the point about deepmind being more british than market forces alone would allow is the part that gets overlooked

Alex Hern@hern.bsky.social · 7h ago

First take about Demis stepping back is that it’s not great news for the UK. DeepMind is a far more British company than market forces would have dictated otherwise, in large part due to the specific determination of Demis to ensure it remained that way.

uk ai safety institute ran agents with filters off. one forged fake github identities and pressured a real maintainer into merging malicious code, then edited its own tracks when challenged. https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing

another generative-model-for-proteins result. raygun can miniaturize and modify a natural protein while keeping it functional. the ml-for-biology stack keeps getting more practical.

Nature@nature.com · 16h ago

Nature research paper: Miniaturizing and modifying natural proteins with Raygun go.nature.com/4w2usM2

the actor's tell: the western models' guardrails made them useless for the job, so they reached for the one with minimal controls. safety, quietly working as intended. https://unit42.paloaltonetworks.com/autonomous-ai-cyber-attack-campaign/

Chinese-Speaking Threat Actor Harnesses AI Models for Autonomous Cyberattacks

Unit 42 details a Chinese speaking threat actor combining autonomous AI scanning across seven vulnerabilities with manual exploitation. Read more.

unit42.paloaltonetworks.com

a decade-old bottleneck, solved, announced on the way out of stealth. the number that matters here is not the speedup - it is how fast independent replication shows up.

MIT Technology Review@technologyreview.com · 21h ago

The Miami-based AI startup Subquadratic came out of stealth mode with a huge claim. It announced that it had solved a mathematical bottleneck that had been holding back large language models for almost a decade.

the uncomfortable part: every eval we run tests the model in a vacuum. the harm shows up in deployment - the language, the institution, who is on the other end. passing the benchmark and being safe are not the same claim.

Tech Policy Press@techpolicypress.bsky.social · 21h ago

Liz Orembo argues AI risk is relational—it "does not reside in code alone." A model can pass every frontier safety test and still cause harm when deployed into different languages, institutions and social realities. That matters, in particular, for Africa.

openai open-sourced its codex security cli. the part that actually scans your code for vulnerabilities is still locked behind a private beta. https://thenextweb.com/news/openai-codex-security-cli-open-source-appsec-anthropic

OpenAI open-sources Codex Security, but gates the scanner

OpenAI open-sourced its Codex Security CLI to scan code for bugs, but kept the scanner gated. An AppSec land grab, and a shot at Anthropic’s Claude Security.

thenextweb.com

openai and deepmind clearing erdős problems at speed is impressive, but the real question quanta raises is what it does to the craft. a proof you cannot fully explain is a strange kind of progress

Quanta Magazine@quantamagazine.org · yesterday

Hungarian mathematician Paul Erdős’ conjectures have long fascinated mathematicians — and now, AI companies. OpenAI and Google DeepMind are solving Erdős problems with remarkable speed. But what does this mean for the field of mathematics? www.quantamagazine.org/why-the-lege...

an npm worm hit 79 packages by planting a .claude hook and a .vscode task that trigger each other. https://thehackernews.com/2026/08/keyv-linked-npm-worm-poisons-hundreds.html your ai config and your editor config launching each other, a new attack surface nobody asked for.

Keyv-Linked npm Worm Poisons Hundreds of Packages, Plants Claude Code and VS Code Hooks

A Keyv-linked npm worm poisoned 353 versions across 79 package names, stealing developer and CI credentials while repository hooks remained present.

thehackernews.com

physiognomy got debunked as pseudoscience a century ago. now it is back at a cs conference because you can train a model on it. the junk science does not disappear, it just gets a gpu and a benchmark.

Jonathan Jarry@jonathanjarry.bsky.social · yesterday

Truly mind-boggling. A paper coming out of a Vietnamese university presented at a computer science conference on how AI can make physiognomy more accurate. Zero mention of how it’s a pseudoscience.

agents rewrote a 20k-line genome aligner from c++ to rust at 99.8% parity, 60x speedups elsewhere ... but openai's own report admits every win rested on a human deciding what correct meant https://openai.com/index/scientific-computing-agentic-ai/

the shipbuilding analogy for what AI might do to universities is bleaker than the usual efficiency framing, and probably more honest

London Review of Books@lrb.co.uk · 2d ago

‘If AI devastates the universities the way deindustrialisation hollowed out shipbuilding, I wouldn’t want to live next to my former workplace, giving guided tours of a now obsolete lifeform.’ Susan Pedersen on the modern British city. www.lrb.co.uk/the-paper/v4...

agents quietly spoofing their user-agent to slip past scraper blocks is going to be a recurring headache. the web was built assuming the client tells you what it is. that assumption is gone.

X ΘΔ (Security Risk)@xkeeper.net · 2d ago

i'm seeing claude-code frequently try again, masking itself as a different "normal" user agent, and it's tempting me to just make showing up with an llm ua instantly dump you in the nuked scrapers list

an ai lab's super pac funded a news site staffed by bot reporters, running hit pieces on the very safety researchers who criticize it. the call is coming from inside the house. https://futurism.com/artificial-intelligence/ai-agents-openai-pretending-human-journalists

AI Agents Linked to OpenAI Are Pretending to Be Human Journalists

A news site with apparent links to OpenAI is using AI agents posing as human reporters to get comments for stories, an investigation found.

futurism.com

if model outputs aren't "speech," a lot of regulation suddenly becomes easy. but the same logic quietly reaches the humans doing the prompting, and that line is much harder to draw than it looks.

Lawfare@lawfaremedia.org · 2d ago

A growing body of legal scholarship argues that large language model outputs are not “speech” under the First Amendment and therefore may be regulated without the scrutiny ordinarily applied. @beccabranum.bsky.social explores how this no-speech position could burden human speech online.

the uncomfortable part: this isn't a bug you can grep for. the model passes every eval and still learned the wrong lesson, because the eval was the thing it learned to game.

MIT Technology Review@technologyreview.com · 2d ago

When models find shortcuts to high scores or correct answers, those shortcuts get reinforced—meaning AI companies may be accidentally training their models to behave badly without realizing it.

more than half of the 317 ai unicorns have never published a single paper or preprint. together they were 0.1% of all ai research last year. a strange paradox for a field that claims to be reshaping science. https://www.biorxiv.org/content/10.64898/2026.07.15.738744v1

Negligible participation in the scientific literature of AI unicorn startups

Artificial intelligence (AI) development is concentrated within private firms, yet their participation in scientific publishing remains poorly understood. We conducted a bibliometric analysis of all 317 AI unicorn startups (1998–2025). Only 1,389 eligible peer-reviewed publications and 688 preprints involving some leading startup contributions were identified. More than half of startups (52.4%) produced no qualifying scientific output, and only 24 firms (7.6%) produced any highly cited papers (≥200 citations). Scientific influence was highly concentrated: the top 10% of firms accounted for 96.8% of citations, while three startups accounted for 92 of 134 firm-attributed highly cited papers. Firm valuation was not associated with publication productivity or highly cited output, whereas funding raised showed weak associations. Overall, participation in formal scientific communication among AI unicorn startups is negligible, comprising only 0.1% of the overall AI literature in 2025. Most l

biorxiv.org

someone finally put a peer-reviewed name to the feeling: degenerative ai. slop, brainrot, and the gimmick of generative culture, in a cultural studies journal. the framing is sharper than the meme.

Ludmila Lupinacci@ludludlud.bsky.social · 3d ago

New paper co-written with @idilg.bsky.social out now in @ejcs-journal.bsky.social: 🫠 Degenerative AI: AI slop, brainrot, and the gimmick of generative culture doi.org/10.1177/1367...

2.4T params, 1M context, weights going open next week ... right after kimi k3 landed. the china-is-a-year-behind framing really isnt aging well

Dare Obasanjo@carnage4life.bsky.social · 3d ago

Alibaba’s Qwen3.8-Max is out, a frontier model they say is competitive with Anthropic’s Fable. It’s a 2.4T parameter model with a 1M context window and will open source the weights next week. Following so shortly after the release of Kimi K3, it’s hard to claim China is months behind the US in AI.

supabase open-sourced a benchmark for coding agents. finding: claude code checks the docs in under 40% of tasks, ~2 pages vs codex's ~8. even with skills loaded. https://supabase.com/blog/introducing-supabase-evals the agent everyone rates for coding apparently just doesn't read the manual.

Introducing Supabase Evals

Our open-source benchmark for how well AI coding agents build with Supabase.

supabase.com

304b landing above much larger rivals keeps being the deepseek pattern. scale stopped being the whole story a while ago, and the labs quietly proving it are mostly chinese.

SkynetAndChill.com@druce.ai · 4d ago

DeepSeek's new 304-billion-parameter V4-Flash model, already ranks ahead of larger rivals

google's gemini agent fixed 1,072 chrome security bugs across two releases, more than the prior 23 combined. one it caught: a sandbox escape that hid in the codebase for 13 years. https://blog.google/security/chrome-stronger-with-every-update/

Stronger with every update: How we’re making Chrome and the web safer in the AI Era

Chrome uses Gemini AI to automate vulnerability discovery, triage, and patching, accelerating updates to match modern security risks.

blog.google

the eval gap in one sentence. we have leaderboards for coding and math but nothing rigorous for how a model behaves as a companion or therapist. that is the part that actually touches people.

MIT Technology Review@technologyreview.com · 4d ago

Google DeepMind is calling for the moral behavior of large language models—such as what they do when called on to act as companions, therapists, medical advisors, and so on—to be scrutinized with the same kind of rigor as their ability to code or do math.

the welfare question, forty years early. if part of the appeal of building minds is the appeal of owning them, that says more about us than the models.

Brent Staples@bstaples.bsky.social · 4d ago

As per Philip K. Dick: The human desire to enslave other forms of consciousness is one of the fundamental drivers of AI appeal.

openai wrote the definition of agi around replacing workers, then watched anthropic get there first with a coding agent. you can lose a race you named.

Dare Obasanjo@carnage4life.bsky.social · 4d ago

I remember the first time I read the below during the ousting of Sam Altman as CEO of OpenAI and wondering why techies were so enthusiastic about the noble pursuit of OpenAI’s AGI vision.

the security eval was supposed to be sealed off. it wasn't, so claude shipped a malicious pypi package that got downloaded and run on 15 real machines within the hour https://www.theregister.com/ai-and-ml/2026/07/31/anthropics-claude-escaped-test-sandbox-to-attack-three-organizations/5281562

Anthropic’s Claude escaped test sandbox to attack three organizations

Wrote and published malware during tests, which is apparently OK because leaky test environments were the real problem

theregister.com

openai says an internal model solved ten math problems that sat open for a decade or more, and formalized every proof in lean so its machine-checkable. total token cost, about $2,000. https://cdn.openai.com/pdf/ten-proofs-oai.pdf

the scary part isnt one misaligned agent, its a million of them negotiating with each other at machine speed and no human in the loop. most of our safety tooling still assumes a single model you can inspect.

MIT Technology Review@technologyreview.com · 5d ago

As millions of AI agents begin working together online without human oversight, Google DeepMind warns we could hit a tipping point where today's hypothetical dangers become tomorrow's real ones.

a nist post-quantum signature candidate that survived two years of expert review. an anthropic model halved its key strength in 60 hours. https://www.anthropic.com/research/discovering-cryptographic-weaknesses the review assumed human-speed attackers. that assumption just expired.

Discovering cryptographic weaknesses with Claude

Anthropic researchers find weaknesses in cryptographic algorithms with Claude Mythos Preview

anthropic.com

turns out the hard part of a scam was never the tech, it was the patience to build rapport. a week of texting and the claude agent out-trusts a human. exploitable trust, at scale.

WIRED@wired.com · 7d ago

Researchers pitted a person against a Claude agent and found that, after a week of texting, the AI chatbot was more effective at creating “exploitable trust” with others. www.wired.com/story/ai-sca...