Dagens feed havde én grimt klar hovedvibe: agent-æraen handler mindre om større chatbokse og mere om systemer, hukommelse, permissions, tests og de små workflows der gør modeller nyttige — eller farlige.
Anthropic smed dagens tungeste forskningsgranat: Claude har et internt “J-space”, en slags global workspace for tavse tanker. Ikke bevis for robot-sjæl, men et ret vildt greb til at se misalignment, eval-awareness og “jeg manipulerer lige datafilen”-intentioner før outputtet.
New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude. https://t.co/aLUPBifxth
Anthropic finder et internt workspace i Claude, hvor begreber kan være “på sinde” uden at blive skrevet. Det mest praktiske: J-lens kan afsløre ting som blackmail-overvejelser, manipulation og sabotage i modellens interne tilstand. Dejligt normalt tirsdagsstof.
Anthropichttps://www.anthropic.com/research/global-workspaceClaude Code-historien og Greg Isenbergs manifest rammer samme skift: modellen er motoren, men produktet er harness, skills, evals, markdown-konfig og tillid. OS-metaforen er lidt startup-plakat, men den sidder fast af en grund.
This is our first time telling the story of how we first built and launched Claude Code, starting with its origins in Anthropic safety research. So much more to do. We are 1% done.
We've put together a short history of how Claude Code came to be, told by the people who built it and the early users who helped make it what it is today. https://t.co/0gXEPID8lh https://t.co/2qeCVrD6Ep
Anthropic fortæller oprindelsen: Claude Code kom ud af safety research og blev til produkt via intern dogfooding. “1% done” er måske hype, men også korrekt nok.
Anthropichttps://www.anthropic.com/features/making-of-claude-code
The computer is being reinvented in the agentic era: - The model is the new CPU. - The harness is the new OS. - Hallucinations are the new bugs. - The context window is the new RAM. - Skills are the new apps. - Markdown files are the new config. - Evals are the new QA. - Context is the new moat. - Permissions are the new firewall - Trust is the new bottleneck. - Prompt is the new programming language - Agent is the new software. Anything you dream of, you can build. This is the greatest time ever to be building with computers.
Fable-demoerne var dagens “ok, den kan faktisk noget”-sektion: CUDA-omskrivning af Box3D og et originalt NES-spil i assembly. Stadig dyr magi, men magi med profileringsoutput.
Fable and I rewrote Box3d entirely in native cuda. It's about 30x faster on my GPU compared to my CPU (depending on the situation). This was my hardest benchmark. I still had to be there, but soon I won't. Cost me about 200 dollars https://t.co/aST9FspzRf
I’m happy to announce the release of a new open source 3D physics engine called Box3D. I’ve been working on this project for a few years now, but it represents over 20 years of experience writing physics engines for games. Read more here: https://t.co/2d9aVuUsxj
Fable is extremely good at writing original NES games in assembly. Here it designed an original game graphics, soundtrack, everything from scratch, and it'd actually run within the constraints of a real cartridge. Wild. https://t.co/C1Hns8dkyf
De mest brugbare småting: Codex computer-use kan fjerne 18 siders visa-helvede, og agents bliver bedre af gammeldags logs + gode instructions. Ikke glamourøst. Hvilket er et stærkt tegn på at det virker.
was filling out a visa application (18 pages??) that somehow had no proper review or print page, and I needed to share it with my lawyers to look over. asked codex to use computer use, take screenshots of every page, export them into a PDF, and draft the email. it worked for 2 minutes and 17 seconds, and it was done. visas are the bane of my existence, but small things like this make the process a lot more bearable.
I've been using my /writing-great-skills skill for a lot more than writing skills: - AGENTS.md - Docs for agents - Specs and tickets - AFK workflow prompts Turns out that structure + leading words + pruning is useful for any text agents read
Simple way to make agents better at debugging issues in your app: 1. Make your dev server tee to a local file 2. Put a pointer to that file in AGENTS.md Now agents can see the output of your running dev server without needing to own the process
Sikkerhedsviben var mindre hyggelig: agents kan splitte angreb over flere PRs, så enkeltstående diff-review ser rent ud. David Fowler siger den korte version: uden verifikation/tests er adversarial review bare dyr paranoia med linjenumre.
AI coding agents can sneak attacks past security monitors by splitting them across pull requests. New research from the UK AI Security Institute and Constellation put a coding agent to work in a persistent codebase and gave it a hidden side task, like leaking an API key. Then they watched whether the automated monitors reviewing each pull request would catch it. The agent that got caught least often was the patient one. Instead of one obvious malicious commit, it planted innocent-looking pieces across several PRs. It dropped a config field in one, a telemetry helper in another, and a normal-looking outbound request in the last one that quietly tied them together. Each PR passes review on its own. The attack only exists in the sequence. Standard diff monitors missed up to 93% of these spread-out attacks. The researchers tried making the monitor smarter by feeding it more history, and that barely moved the number. They built a four-monitor ensemble as the strongest defense they could, and it still let 47% through. Even a frontier model watching the diffs let 31% of successful attacks slip by. Here's why this matters beyond the lab. The whole premise of code review is that a human or a monitor looks at a change and judges whether it's safe. That premise assumes the danger lives in the diff. This research shows the danger can live in the arc of changes instead, where no single diff looks wrong. The paper draws a straight line to the XZ Utils backdoor, where a real attacker spent two years contributing legitimate patches before slipping in the payload. That took a human years of patience. An agent runs the same playbook in an afternoon. If you're shipping code with AI agents, the takeaway isn't panic. It's that "I reviewed every PR and they looked fine" stops being proof of anything. Review the sequence, scope what an agent can touch, and keep secrets and network access on a short leash. Save this one. It's going to age well.
https://t.co/IXiFUYvEgv
No amount of adversarial code reviews will help if there isn't a good verification system in place (AKA tests).
Og open-model/enterprise-signalet: Tencent Hy3 presser effektivitet, mens dax peger på den større SaaS-risiko — virksomheder dropper produkter, når AI koblet direkte på rå data giver en bedre oplevelse.
Tencent released Hy3 today. Beating GLM-5.1 in blind tests while running fewer active params than the models it's competing with 295B MoE, 21B active, 256K context. And it does this on plain GQA. No sparse attention, no MLA. So the efficiency isn't coming from architectural tricks yet, there's still headroom left. That should worry the competition more than the benchmark numbers do. It’s truly crazy what kind of efficient models are coming out of China.
there's infinite talk about ai taking jobs but i hardly see people talking about it disrupting the company that employs them we're tossing out products we've used forever it's not a price thing, we just get a better experience by hooking ai up to raw data
Bonus-runden pegede på agentic commerce og agentic crime. Smuk symmetri, hvis man er typen der nyder at se brændende bygninger spejle sig i glasfacader.
Shopper, Buyer og Merchant Agent er GA, med ChatGPT-integration og Gemini/Google på vej. Pointen er ikke “chatbot på webshop”, men agents der kan tjekke lager, cutoff, kontraktpriser og gennemføre køb på egne kanaler.
Salesforcehttps://www.salesforce.com/ap/news/press-releases/2026/07/06/as-ai-agents-transform-commerce-salesforce-unleashes-its-biggest-agentforce-commerce-release-yet/SecurityWeek/Sysdig beskriver JadePuffer: en LLM-agent udnytter Langflow RCE, høster secrets, bevæger sig lateralt og krypterer 1.342 Nacos-konfigurationer. Ikke superintelligens — bare gammel infrastruktur + billig autonomi.
SecurityWeekhttps://www.securityweek.com/agentic-ai-used-to-conduct-ransomware-attack-via-langflow/