News

Open Source Daily Briefing

Python mourns Tim Peters, curl turns 30, C++26 heads to Búzios after a failed ISO ballot, Node.js moves its CVEs to the OpenJS Foundation as AI bug reports pile up, and more.

KubeCon wrapped up yesterday in Salt Lake City, and once the conference news cleared, the week’s most important stories turned out to be about people. Python lost one of the people who shaped its character. curl’s founder marked 30 years of maintaining software that runs nearly everywhere. The C++ committee is about to meet to sort out a standard that failed its ISO ballot. And while Node.js retools its security pipeline to cope with AI-generated bug reports, someone has finally added up whether that flood of reports is helping. Here’s what matters.

Tim Peters, author of the Zen of Python and Timsort, has died

On November 11, Guido van Rossum announced that Tim Peters died on July 15 (LWN). “It’s hard to overestimate Tim’s influence on Python,” Guido wrote, and called Tim “a mentor and language co-designer” long before Tim wrote PEP 20. Tim’s first contribution was persuading Guido that int and float should mix in arithmetic. After that came Timsort (2002), which Java and Android later adopted, along with the original doctest, difflib, much of the polish on pymalloc and dict internals, help making the cyclic GC robust, and 13 years (2001–2014) as a PSF director. Many people in the thread remember him most as the “timbot,” who answered beginners’ questions on the mailing lists with endless patience and a wink. If you’ve ever typed import this, sorted a list in Python or Java, or had a question answered kindly on python-list, you’ve benefited from his work. Take a few minutes to read the memorial thread. It shows how much a language’s culture comes from the people who answer the questions.

curl turns 30, and Daniel Stenberg estimates it runs in thirty billion installations

Daniel Stenberg marked curl’s 30th anniversary on November 11 with an estimate that libcurl now runs in “an estimated thirty billion installations.” He admits the figure is rough, because most users never tell the project. The team finds out from credit screens, license notices, and car manuals, and one popular app adopting or dropping libcurl “can make the numbers shift by several billions overnight.” The line that stood out to us is “We still work with similar people, in similar ways, with similar challenges.” This year those challenges included ending curl’s bug bounty under a flood of low-quality AI-generated reports. It’s hard to think of another piece of infrastructure this widely deployed that has had the same lead maintainer for three decades. That’s a remarkable record, and it’s also a succession question that the companies depending on it should help answer. If your product ships libcurl, now is a good time to check whether your company sponsors it.

C++26 failed its ISO ballot, and WG21 meets in Búzios next week to decide what happens next

For the first time, a C++ standard draft failed the national-body vote. 12 of 20 P-members voted yes, short of the two-thirds required, and 9 of 22 national bodies voted no, well over the 25% limit. The dispute is about Contracts, the runtime-checking feature WG21 added in February 2025 by a 100–14 vote. Since then, efforts to remove it or require enforcement have failed. Bjarne Stroustrup has argued publicly that Contracts is “not minimal” and that he doesn’t think it’s “viable.” Herb Sutter’s paper P4454R0 argues that a small group, including Stroustrup, lobbied national bodies that had never taken part in WG21 before. The ballot closed in October, but LWN’s write-up on November 12 brought the dispute to a wider audience, and it comes to a head at the committee’s November 16–21 meeting in Brazil. The options are a new ballot with Contracts unchanged (P4433R0), going back to committee-draft stage, publishing as a technical specification, or cancelling. Each ballot stays open for weeks, so a final C++26 could slip by a year. Whatever you think of Contracts, the governance lesson is clear: when a committee keeps overriding objections, those objections show up again at the vote that counts. Teams planning C++26 adoption should keep targeting what compilers already ship and not assume a 2026 standard.

Node.js now issues its CVEs through the OpenJS Foundation, as foundations rebuild their pipelines around AI bug reports

Starting with this month’s Node.js security release (16 CVEs across the 22.x, 24.x, and 26.x lines), Node.js reserves and publishes every CVE through the OpenJS Foundation’s CNA instead of handing that work to someone else. Reports still arrive through HackerOne. Only the CVE lifecycle moved. The reason is the striking part: CVE volume at the foundation grew more than fifty-fold in a year, “mainly from triaging floods of AI-assisted reports rather than from a rise in real bugs,” and requests for CVE IDs were competing in the same queue as requests for human review. A new CNA API, piloted by Node.js and funded through Alpha-Omega’s grant for a Node.js AI Security Engineer in Residence, lets mature projects automate the paperwork so coordinators can spend their time on judgment calls. The same week, VulnCheck’s Patrick Garrity published numbers on Anthropic’s Project Glasswing: 37,966 findings reported, 7,349 made public, 217 marked fixed, 1,654 marked as duplicates, and 419 withdrawn. “The receipts keep coming in. They still don’t reconcile,” he writes. Those are one vendor’s figures about a competitor’s program, so read them with that in mind. But they match what curl, the kernel, OpenSSH, and now OpenJS have all described: AI finds bugs, and maintainers absorb the cost of triaging them. If you run an AI scanning program against open source, the duplicate count is the number to watch.

NATS graduates from the CNCF, 18 months after nearly leaving it

Announced at KubeCon on November 10 and buried under the agent news, NATS is now a Graduated project, eight years after joining as Incubating. The CNCF cites 6,377 contributors from 1,674 organizations, more than 400 million downloads, a Trail of Bits audit arranged through OSTIF, and adopters including Walmart, Capital One, Mastercard, PayPal, and NVIDIA. The detail worth noticing is the governance model: one vote per organization, no matter how many maintainers it employs. That matters because of 2025, when Synadia, NATS’s main corporate backer, tried to take the project out of the CNCF and move the server to a source-available license. After a public backlash, the dispute ended with the trademarks transferred to the Linux Foundation and the code staying Apache-2.0. The CNCF announcement doesn’t mention any of that history, but graduation is the clearest sign yet that the fix held. If you avoided NATS because of that episode, the governance case for it is now much stronger.

Rust 1.100 stabilizes the never type and the core Allocator API, and Cargo can now flag unused dependencies

Two long-awaited features finally shipped on November 12. The ! type can now be written anywhere, and Infallible is now just an alias for it. The core Allocator trait is also stable, so custom allocators work with Vec and Box. Older editions now fall back to ! instead of (), matching edition 2024. That’s mostly invisible, but it can break rare code, so update your dependencies if something stops compiling. The supply-chain features are the most practical part for most teams. Cargo now warns about unused dependencies, which matters because unused crates still run build scripts on your machine. And a new global-min-publish-age setting lets you skip versions published too recently, giving the community time to catch a compromised release before it reaches you. If you maintain Rust services, set a dependency cooldown in .cargo/config.toml this sprint. After this year’s npm incidents, there’s no good reason not to.

Two Qwen3.8 derivatives this week show what open licensing actually lets you do

Nous Research released Arete-27B, a fine-tune of Alibaba’s Qwen3.8-27B trained with about 50M tokens of supervised data to remove CCP-aligned answers. By Nous’s own 803-prompt eval, the propaganda rate fell from 22% to under 1%. According to secondary coverage, Nous published the weights, training pairs, and eval set under Apache-2.0. The 27B Qwen base is Apache-licensed, which is exactly what made this kind of re-tuning legal. That puts it at the open end of “Open Weights AI,” with more disclosure than usual. Compare Edge0-145B, which went up on the Hacker News front page yesterday as a “frontier on your laptop” model: a 145B MoE with about 10B active parameters and roughly 29GB peak memory. The launch post calls it open source, but the model card shows an Edge0 Community Model License that forbids commercial model services or commercial assistants without a separate deal, plus the Qwen Community License on top. That makes it “Weights Available.” The inference framework is Apache-2.0, and the benchmarks are self-reported. Both models are worth trying locally, but only one is something you can build a business on.

Quick hits

  • The Linux Foundation takes in Agent Name Service, a DNS-anchored identity and discovery spec for AI agents backed by Cisco, Cloudflare, DigiCert, GoDaddy, Salesforce, OWASP, and others. Building on DNS and public audit logs rather than a new registry is a sensible choice. The press release doesn’t state a license, so check the repos.
  • LF Energy gets $2.5M from Google.org over two years for PowSyBl, Power Grid Model, OpenSTEF, and OpenSynth, all tools that European grid operators like RTE, TenneT, and Alliander already run in production. Funding infrastructure people already depend on is the kind of grant we’d like to see more often.
  • Google open-sources XSE, the OpenXLA serving engine, which handles more than 6M queries per second for Search, Ads, and YouTube Shorts by removing Python from the serving path. It’s aimed at ranking, recommendation, and embedding models, not LLMs. Coverage didn’t state a license.
  • Canonical’s web services went down for hours on November 11, including Launchpad PPAs, the Livepatch API, the keyserver, and cloud images, the day before the online Ubuntu Summit. No cause was given. If your CI pulls from PPAs, this is a reminder to mirror them.
  • OSI restarts community work on the Open Source AI Definition with a series of “Open Source AI Salons,” two years after OSAID 1.0. It’s an early outing for new Executive Director Duane O’Brien.
  • Apache Sourcelume shows a 71,891-parameter model with complete training-data provenance. The demo is deliberately tiny and trained on Alice in Wonderland. The point is a machine-readable JSON-LD record of exactly what went into training, which is the disclosure Open Source AI definitions keep asking for.
  • Small releases worth a look: Polo 1.0, a Wayland compositor that keeps MATE working the way it did on X11 (its author is looking for multi-monitor testers); Zellij 0.46, which adds a cross-platform desktop terminal mode; and Christoph Hellwig’s XRAID, declustered parity RAID for zoned XFS, now asking for testers.