After AI scrubbed away the provenance, an open-source project switched its license from MIT to GPL

Today is Friday, November 20, 2026. This week, two things in the open-source community seem unrelated, but they are actually about the same thing.

One came last week—Asahi Linux announced that it was moving all repositories from MIT to GPL-2.0-or-later, not because of legal risk or commercial calculation, but because "someone had laundered its code with AI." The other came a few days earlier—a paper team used a large model to scan more than 4,000 C/C++ projects and confirmed 187 memory-safety defects; almost simultaneously, another third-party bill showed that a certain big tech company's AI vulnerability mining project had reported 37,966 "findings" in seven months, with only 217 marked as fixed.

Putting them together, one signal becomes clear: AI pushes the cost of "production" down to near zero, but the cost of "acceptance" still rests firmly on people. And the vast majority of ledgers only count the former.

1. A decades-old “signature” rule has been broken by detection costs

First, Asahi Linux. It is the underlying project that lets Linux run on Apple Silicon Macs. For years, it has placed most of its code under the MIT license, precisely so that anyone can reuse it as needed. Now that road has come to an end.

The project states it plainly in “Progress Report: Linux 7.3” (James Calligeros, 2026-11-07): they have repeatedly found people using permissive licensing to hijack the work, with the goal directly aimed at “replacing this project and aggressively monetizing it on Linux on Apple Silicon.” In some extreme cases, feature code that is still in development and has not yet been officially released has been taken directly from personal forks and published without attribution, or rewritten with a large model to conceal its origin.

This is the so-called “code laundering”: taking someone else’s code, running it through a process to wash its provenance clean, and then using it without complying with the original license or giving attribution. Before AI emerged, this process required changing it line by line, renaming variables, and moving functions around; large models have made it cheap and fast.

The MIT license only requires preserving the original copyright notice; it does not require disclosing modifications or making your version continue to be open source. This is both why it is popular and exactly its weak point. In the project's own words: MIT "explicitly permits this behavior, and has no terms that can guard against malicious actors." So Asahi's decision is: from now on, all code will use only strong copyleft, namely GPL-2.0-or-later.

This is not the product of a legal department, but an institutional choice forced upon a project that had been pushed into a corner: when “attribution,” the oldest informal constraint, was breached on detection costs, the only thing left to bet on was an enforceable formal institution.

The migration is still underway and did not happen overnight. The verifiable supporting evidence on GitHub is: AsahiLinux/m1n1's license metadata is already GPL-2.0, AsahiLinux/asahi-nvram is still MIT, and AsahiLinux/linux is marked as NOASSERTION—a batch of repositories is in an intermediate state.

A reminder for downstream users of the enterprise: Repositories whose license is "in transition" are exactly where compliance judgments are most prone to error—the code you pull today is MIT, and tomorrow an update to the same repository may already be GPL-2.0. Dependency locking cannot lock only the version hash; it must also lock the "license snapshot at the moment of retrieval."

II. WINE already took this path twenty years ago

Asahi itself compares this decision to WINE: WINE was also initially MIT, and from the 1990s to the 2000s it was harvested by a large number of proprietary forks; in 2002 it was changed to LGPL, in order to block further proprietary forks and encourage everyone to contribute directly upstream.

History rhymes a little ironically. Twenty years ago, what was guarded against was "manually modifying code to create a closed-source fork"; twenty years later, what is guarded against is "AI laundering code to create a substitute product." What has changed is only the cost of laundering; what has not changed is the inherent powerlessness of permissive licenses against "free riding."

WINE · 2002Asahi · 2026
MIT → LGPL, addressing proprietary fork harvesting, encouraging upstream contributionsMIT → GPL-2.0-or-later, addressing AI content laundering and commercial free-riding

What is truly worth noting is that Asahi did another thing in the same year: it changed some drivers (TI USB-PD, SPMI, initial Thunderbolt support) to upstream-first development, and for the first time sent Thunderbolt patches directly upstream.

Tightening licensing while improving upstream collaboration—these two things happened in the same project and the same year; it is no coincidence. In institutional economics terms: boundary rules (who can take it, how it can be used) and cost-sharing rules (who maintains it, who contributes) must change at the same time; if only one is changed, the cost will be pushed elsewhere. Use copyleft to block free riding, and upstream-first to raise the substitution cost of "going around you."

III. "Discovery" Is Now Free, but "Acceptance" Has Become a New Scarce Commodity

The Asahi story is 'code laundering,' while another set of data from the same week says 'report padding'—the two are two sides of the same disease.

According to a review of the 11-15 briefing from “The Way of Open Source,” VulnCheck tracked the seven-month ledger of a major company’s AI vulnerability discovery project: as of 11-02, it self-reported 37,966 findings, only 7,349 (19.4%) were made public, and only 217 were marked as fixed. In the same period, findings increased by 11,813, while fixes increased by only 15.

55.9%

The proportion of the 34 AI vulnerability reports publicly disclosed by curl that were judged “not a vulnerability”

The curl project received 45 reports; of the 34 that have been made public, 19 were judged "not a vulnerability" (55.9%), only 17.6% were confirmed as vulnerabilities, and all of them were low. Another paper turned "AI junk PRs" into a project-level empirical study for the first time: among PRs that maintainers deemed low-quality AI contributions, only 14% of authors proactively disclosed using AI; the vast majority were closed within three days, and when closing them, maintainers often gave no reason at all.

segmentBefore AI emergedAfter AI emergedWho bears the cost
Output Code / ReportExpensive, scarceApproaching zeroSubmitting party, almost zero
Review / AcceptanceMatches outputBecome the new bottleneckMaintainers, enterprises
Attribution / Source EvidenceRely on voluntary complianceTesting costs have been breachedOriginal author, legal

These numbers remind us of one thing: any governance design premised on "requiring disclosure of AI use" must first look at that 14% disclosure rate. If the system still assumes "high submission costs," the screening burden will fall entirely on the gatekeepers—this is institutional lag, not a moral issue.

IV. What CIOs should do is not use less AI, but keep acceptance authority in the enterprise's hands

For those responsible for enterprise digitalization, the crux of this matter is not "whether to use open source," but "when using open source, on what basis can we trust it, and how can we prove that we haven't crossed the line."

1. A license is not a one-time legal signature; it is a dynamic field that must go into the ledger. Changes like Asahi moving from MIT to GPL-2.0 affect downstream distribution and commercial terms. An enterprise-level dependency list should record "license + version + evidence date + source of change notification," rather than stopping at the PDF version from the time of procurement.

2. Put "signature verification" and "provenance" on the acceptance checklist.Critical open-source dependencies must be able to answer three questions: where artifacts come from, whether builds are reproducible, and whether the provenance chain is complete. SBOMs, artifact proxies, and signature verification—these three—have gone from "nice-to-have" to "must-have for survival" after the cost of code laundering fell to zero.

3. For AI code generated internally, establish ownership rules before discussing efficiency. When enterprises use large models themselves to write code and generate patches, they likewise create the problem of "unclear provenance": which open-source code was this code modified from, whether it should inherit the original license, and whether external distribution is permitted. Writing these into AI coding standards is much cheaper than assigning responsibility after the fact.

Go back to those two sets of numbers at the beginning. They seem to be saying different things—one project changed its license, and a batch of reports was ruled invalid. But they point to the same judgment:

When generation is nearly free, value shifts from "whether it can be built" to "whether it can be clearly verified, clearly accounted for, and clearly traced to its source." Enterprise open-source governance and AI governance must ultimately all come down to the same set of auditable ledgers.

This also explains why over the past week, whether it was issuing “non-human identities” to agents, or arguing over who has the right to define “what counts as a defect,” everyone was in fact answering the same question: in a world where everyone can produce at low cost, who is accountable for the results, and on what basis should people trust them.


Primary sources and dates: "The Way of Open Source" 2026-11-15 briefing (including Asahi Linux《Progress Report: Linux 7.3》, James Calligeros, 2026-11-07; LWN.net 2026-11-07 excerpt; VulnCheck 2026-11-09; papers such as arXiv:2611.15159, arXiv:2611.16565). The relevant figures are all based on public reports and paper statements; some come from third-party tracking or researchers' self-reports and have not been audited item by item. When citing, please refer back to the original texts.

This article was compiled by Zhongdi Information for the observation of enterprise digitalization and open-source governance.

关于我们

​我们致力于帮助中小企业实现数字化转型,我们的团队由一群充满激情和创新思维的专业人士组成,他们具备丰富的行业经验和技术专长。

扫一扫获取顾问以及手册

Sign in to leave a comment
You didn't import it, yet it stole your cloud keys: a poisoning attack targeting an AI agent development framework