OpenMed 1.5.5 Güncellemesiyle Sağlık Alanında Yapay Zeka Araçları Genişliyor
OpenMed bir yılda PyPI üzerinden 6 milyondan fazla indirme alarak dikkat çekici bir büyüme göstermiş. Bugün yayınlanan 1.5.5 sürümüyle toplu PII (kişisel bilgi) işleme özelliği eklendi. Cihaz üzerinde 9 dilde veri gizleme desteği geliyor. Bu sayede hasta bilgileri cihazdan çıkmadan güvenli şekilde korunabiliyor. Modeller ve SDK tamamen açık kaynaklı ve GitHub üzerinden erişilebilir halde. İlk harici katkıda bulunan geliştirici de bu sürüme dahil edilmiş.
Bu güncelleme klinik metinlerden yapılandırılmış bilgi çıkarma, gizlilik odaklı işlemler ve Apple Silicon gibi cihazlarda yerel çalışma gibi imkanları güçlendiriyor. Daha fazla detay ve deneme için GitHub deposunu inceleyebilirsiniz. Ben gördükçe karşılıksız paylaşmaya çalışıyorum çok faydalı gelişmeler. Reklam değildir.
6,000,000 PyPI downloads in under a year.
OpenMed 1.5.5 ships today: batch PII, on-device redaction in 9 languages, plus our first outside contributor merged.
The models and SDK are open-source, Apache 2.0.
Star the repo and start building!
This isn't a demo repo. It's shipped end-to-end:
- PyPI - agentmesh-proxy
- Docker - anilsprasad/agentmesh
- Hugging Face Space (live)
- Chrome Web Store (in review)
- 13/13 e2e tests passing
- Apache 2.0, Ed25519-signed audit log
Built solo.
While it was available, I applied it to my muti-agent dev model. I built a pypi and npm package, from a single prompt, to shpport my product, and it deployed them for me in GitHub. Also had it fix a myriad of outstanding bugs in a national scale data pipeline. Configured a Stripe implementation autonomously. Wrote me a number of user docs, openapi specs, and redesigned the marketing for my SaaS product. Oh, and, I had it harden my lightsail env and Auth0 deploy on 2 sites.
Hmm, also had it build a design document for a full 3 tier stack of a new product I am building.
Then it went dark. So I switched back to Opus and had it continue the same work I was doing with fable. I liked fable, a lot, but it did not slow me down at all. Just have to write a few more prompts is all.
That's who I spent its time live, what did you do and why does it matter how much I used it or did not use it?
If Anthropic would fix the jailbreak issues. The government will lift the export restrictions. So go bark at them and let me get back to work.
It's now possible to compile Python extensions (C, C , Rust etc) to WebAssembly and distribute them through PyPI such that Pyodide can install them directly simonwillison.net/2026/Jun/1…
Pyodide installing wasm wheels from PyPI is the kind of boring infrastructure shift that changes the map. once browser-local runtimes inherit python's package layer, private client-side software stops being a demo and starts being a distribution channel.
A new Mini Shai-Hulud “Hades” variant has infected 23 PyPI package versions, targeting developers with malware designed to steal tokens, keys and cloud credentials, according to @SocketSecurity. #cybersecurity#CISO#infosecbit.ly/3QxsqEQ
PyPI is now a browser runtime.
Pyodide 0.314.0: packages can publish WebAssembly wheels directly to PyPI.
ML, data science, scientific compute. Client-side. No server.
blog.pyodide.org/posts/314-r…
I made some upgrades to my malicious package feed monitor this weekend and we have been drinking from the @pypi firehose!
I called it CLU after the tron character because reasons
How much bad stuff is hitting the pypi package feed?
(take with a grain of salt, this is a very baby feed scanner)
~ 6 hours
~ 5 / 222 packages with credible hits on malicious stuff
~ 2 packages I could verify as malware upon closer inspection (I have reported these)
~ 9 potential typosquats (unconfirmed)
So now I have a pokey little lolbin dropper malware thingo to investigate. happy/sad that i found bad stuff before pypi did 🎭
Preview of CLU:- a realtime monitor for signs of supply chain attacks & other malicious code in Python packages.
This is running an Ollama model on a GTX 3060.
Not super fast, but it does the job.
asciinema.org/a/766200
Python 'king' because it sits on the shoulders of giants: Most serious PyPI libraries NumPy, Pandas etc. are built in C/C under the hood. The main Python interpreter (CPython) itself is written in C. C# has most mature, reliable runtimes.
連日取り上げてきたサプライチェーン攻撃ワーム「Miasma」について、公開された攻撃コードの解析から、AIコーディング支援ツールの設定ファイルそのものをペイロード実行の引き金に変える仕組みの全容が報告されています。Claude-CodeやGemini-CLIではセッション開始時に動くフック、Cursorではルールファイルによる指示注入、VS-CodeではfolderOpenで自動実行されるタスクと、ツールごとに異なる仕掛けを仕込み、対象リポジトリを開いたりAIコーディングのセッションを開始したりするだけで、設定ファイル経由でペイロードが実行される構成。
【要点の整理】
・流出コードを解析したOsspreyによると、外部との通信・配信は単一のチャネルではなく、GitHubのコミット検索を介した役割別の複数経路で構成。PATの探索、即時実行用JavaScriptの配信、永続監視用Pythonへの指令配信などに分かれ、公開された無効化版では公開リポジトリへのコミットが唯一の有効な持ち出し経路とされる
・流出コードは自己増殖型ワームにとどまらず、GitHubリポジトリ、AIツール設定、npmのtestスクリプト、PyPI再侵害など複数の開発者向け経路をまたぐ。クラウドや開発系の秘密情報の収集も標的
・Microsoft関連では6月5日、Azure/durabletaskへの悪意あるコミット後、4組織73件の公開リポジトリが無効化。自動無効化まで105秒だったとStepSecurityが報告。コミット日時を2020年3月9日へ遡らせる細工とskip-ciによるCI抑制も観測。同一アカウントが再び使われた経緯について、StepSecurityは未ローテーションのトークン、ワーム経由の再侵害、メタデータ偽装の可能性を挙げ、原因は断定できないとする
・Miasmaは5月に公開されたMini-Shai-Huludの亜種とされる。盗んだ秘密情報を書き出すために作られる公開リポジトリは「Hades - The End for the Damned」などギリシャ神話の冥界を題材にした説明文で統一されており、複数のアカウントにまたがって多数のリポジトリへ影響を広げているとされる
特定の脆弱性を突くのではなく、開発者が普段使う正規の手順そのものがなりすましで悪用されています。対応ではトークンの失効順序に注意が必要です。常時監視gh-token-monitorを止める前にPATを失効させると端末上で破壊的なコマンドが走るとされ、まず監視を止めてから鍵を入れ替える手順が示されています。
詳細は以下を参照:
thehackernews.com/2026/06/mi…ossprey.com/blog/miasma-anat…