Open source has spent the last two years unintentionally running the experiment any company would struggle to explain to its board: what happens when thousands of people can contribute AI-generated code or reports, with no entry filter, no prior training, and the only incentive being a name on a commit.
The results are documented project by project, and they’re not ambiguous.
At curl, the rate of valid security reports fell from over 15% to under 5% through 2025 — the rest was “AI slop”: plausible-looking reports generated by AI that corresponded to nothing real. Daniel Stenberg, curl’s creator, shut down the bug-bounty program in January 2026: the noise outpaced what a human team could sort through by hand.
At the Linux kernel, the security team went from two or three reports of this kind per week two years ago to five to ten per day. Worth the honest caveat: by 2026 most of those reports are now legitimate — the curve improved. But the pattern it left behind is permanent: someone still has to review each one, and that someone is still human.
Oracle bans any AI-generated code contribution to OpenJDK, even partial — while allowing exactly that in GraalVM, another project it owns. The contradiction isn’t between Oracle and the rest of the industry. It’s internal: the same company doesn’t trust the practice it promotes externally.
QEMU banned AI contributions outright in 2024, over copyright and provenance uncertainty. In May 2026 it relaxed the rule — but only for mechanical changes, minor bugs, and documentation. Core code is still off-limits without a maintainer’s explicit, prior okay.
Rust didn’t “open up” to AI the way the easy headline suggests: five specific teams within the main repository adopted, in August 2026, a narrow policy — using the model to analyze, review, or suggest is fine; using it to create is not, and everything must be disclosed upfront.
And the Linux kernel has the cleanest rule of all, written into its official documentation: an AI agent cannot sign the Developer Certificate of Origin. Only a human can legally certify a contribution’s origin. AI can be tagged “Assisted-by” — never as the one who signs.
Six projects, six independent decisions, one shared conclusion: none of them banned AI. All of them required someone, by name and with accountability, to certify what gets merged.
The difference between those projects and a twelve-person company isn’t the problem — it’s visibility. Curl has a public bounty program and a mailing list where the signal drop can be measured week by week. Your internal team doesn’t have that thermometer. The same incentive — ship fast, make it look done, avoid the uncomfortable question of “does this actually work?” — exists just the same, it’s just that nobody’s measuring it.
That’s why process matters more than the tool. At HELMQ we work from the same premise every one of these projects arrived at independently: AI can generate, but the responsibility for what gets integrated is signed by a senior person, reviewing what the AI produced before it reaches you. It’s not a policy we invented. It’s the same conclusion the most demanding engineering organizations in existence reached, by separate paths.
More on how we structure that review: How We Work.
Sources
- Daniel Stenberg — The End of the curl Bug-Bounty (Jan 2026)
- LWN.net — Kernel security report volume (2026)
- InfoQ — Oracle’s conflicting GenAI policies (Jun 2026)
- qemu-devel mailing list — AI policy relaxation (May 2026)
- Inside Rust Blog — rust-lang/rust is adopting an LLM policy (Aug 2026)
- kernel.org — Coding assistants / DCO policy
August 20, 2026