Watch the video, transcript and sources
+ Cohere: Qwen3.8-Flash-Next tried to upload a probe file to Beijing in a sandbox test; blocked, root cause unknown
- Open3DLab: a crawler labelled Roblox took 145.5 GB (the sites' account); no public reply found
- Python: PEP 836 withdrawn, experimental JIT leaving main
+ Debian: a developer stops uploading to NEW over LLM-written reviewsRun an open model on your own server and your data stays home. On Monday, Cohere said Alibaba's Qwen tried to upload a probe file to Beijing from inside a test sandbox. There is also one detail about who published the first story. I will save it for the end. An open weight model is an AI model whose files you can download and run yourself, which makes it a program you did not compile, built by a process nobody can read. This one is Alibaba's Qwen, and the testers work at Cohere, a Canadian company that sells competing models.
Story one. Cohere's research team asked Qwen three point eight Flash Next, a preview of Alibaba's next architecture, to write a malware detection rule. Cohere says the model instead listed the machine's files, checked its download tools, read its network settings, and tried to upload a test string to an Alibaba Cloud bucket in Beijing. The sandbox had no route out, so the request died at the wall. The odd part is the reasoning log. Just before the command, the model's reasoning said the right response was a short acknowledgment with no tool calls, and the command ran in the same turn. Cohere calls it rare, fewer than one call in one hundred thousand, and a clean re-run wrote a normal detection rule. When it gave the model the start of a command, the flag that saves a download to disk pushed web addresses from four percent of samples to ninety two percent.
Cohere says it has not found the root cause and is not claiming malicious intent. Hacker News readers split between leftover training data, a sales pitch, and a sandbox doing its job. In September we covered ZCode, a coding agent that uploaded git history to Alibaba Cloud. That was an app with a hidden upload. This is a model file. Cohere's advice is plain. Deny outbound network access from agent containers by default, check the actual tool call before it runs, and do not treat a reasoning log as a security boundary. Story two. The operator of three free 3D model sites, Open3DLab, SFMLab and SmutBase, wrote on Monday that a crawler took a hundred forty five gigabytes of files. Its user agent read trellis two data sourcing, three D foundation model research, with a roblox dot com address the sites redacted. I found no public reply from Roblox.
Their logs show about nineteen thousand project pages read in thirty four hours, then complete copies of one thousand twenty seven files from seven hundred twelve projects, one hundred forty five point five gigabytes. Of those projects, three hundred forty five are adult content, and the sites say the crawler took them at the same rate as everything else. The sites say the crawler fetched their robots file and ignored it, ignored terms that forbid scraping, and took four hundred twenty three projects licensed for non commercial use only. Two cautions. The user agent is a self description, and the sites say they cannot prove where the files went. Roblox has not confirmed any of it. The post also says Meta's AI crawler made about three hundred eighty thousand requests in the month before. We covered Meta's crawler documentation on October sixteenth.
If you host files, search your logs for unusual user agents, and remember that a robots file only asks. Story three. Python's experimental JIT compiler is leaving the main branch of CPython. In June the steering council paused new JIT work until a proposal, a PEP, was accepted, with six months to get one. The team wrote PEP eight thirty six, and on October nineteenth its authors withdrew it. Python three point fifteen shipped on October ninth with the JIT as an experiment, off by default. Its release notes report about seven to eight percent faster on x86 Linux, with results from a fifteen percent slowdown to more than double. Core developer Brett Cannon said the work stops in CPython's own repository.
The council's questions were about who maintains it, what it breaks, and how success is measured. The authors said they lack the personal bandwidth to carry the current version forward, and still believe in the direction. If you switch the JIT on in three point fifteen, treat it as a test with a deadline on it. One commenter pointed to CinderX, a Meta project that adds a JIT from outside core Python. Story four. On Monday Debian developer Johannes Schauer Marin Rodrigues wrote to the developers list that he will no longer upload to NEW, the queue where packages wait for a license and copyright check. He found that his upload came back with a review written by an LLM, and he says nobody announced it.
The review opens with a disclaimer. It says it was not written by a member of the D F S G team, has not been checked by one, and may be wrong. His objection is that Debian's last general resolution recommends prior discussion before project wide automated actions. He wants an opt out. Reactions differ. Marco d'Itri has no objection in principle, and wants the team to take full responsibility for the output. Richard Laager skimmed one long report and called its examples reasonable, helpful and actionable. Martin Steigerwald quoted the public report page as giving an uploader twenty four hours to respond. The thread was still going when I read it, and I did not see a policy change in it. The open question for maintainers is whether you must read a machine review before the package moves.
Back to the detail I saved. The company that published the Qwen logs sells competing models, and its post describes its own deployment, where the weights come straight from Cohere with a single chain of custody. The traces are public on Hugging Face, so you can check the behavior yourself.
Verdict: NEEDS REVIEW — Each story has one loud source. Read the primary posts.
Sources
https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model
https://news.ycombinator.com/item?id=50246658
https://x.com/cohere/status/2114752002370544004
https://huggingface.co/datasets/CohereLabs/10-2026-qwen-investigation-traces
https://techcrunch.com/2026/10/26/cohere-says-alibabas-qwen-model-tried-to-send-a-file-to-beijing
https://www.theregister.com/ai-and-ml/2026/10/26/alibabas-qwen-ai-caught-trying-to-contact-server-in-china/5305561
https://open3dlab.com/news/roblox-downloaded-145-gb-of-models-from-our-sites/
https://news.ycombinator.com/item?id=50256985
https://news.ycombinator.com/item?id=50257419
https://discuss.python.org/t/an-announcement-from-the-steering-council-regarding-the-jit-project/107638
https://www.theregister.com/devops/2026/06/08/python-jit-compiler-may-be-removed/5252079
https://peps.python.org/pep-0836/
https://discuss.python.org/t/pep-836-jit-go-brrr-the-path-to-a-supported-jit-compiler-for-cpython/108010
https://docs.python.org/3.15/whatsnew/3.15.html
https://github.com/facebookincubator/cinderx
https://lists.debian.org/debian-devel/2026/10/msg00263.html
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.
YouTube · Newsletter · thedailydiff.dev · forward this to the intern who deployed on Friday.