Open-source projects thrive on outsiders reporting bugs. Yet this very mechanism is currently under pressure: the maintainers of Curl – the widely used command-line and library building block for data transfer via HTTP and other protocols – report that they can no longer cope with the flood of incoming AI-generated bug reports. To protect themselves, the project is now resorting to drastic measures.
The wording from within the project is clear: action is needed in order to survive. That sounds dramatic, but it describes a sober capacity problem. Maintainer time is the scarcest resource in many open-source projects, and it does not scale with the number of incoming reports.
Why automated reports are particularly costly
A bug report costs a project effort even if it ultimately turns out to be unfounded. Someone has to read it, trace the described code path, reproduce it if necessary and reject it with reasons. In the case of security-relevant reports, there is the added factor that they must be treated confidentially and cannot simply be discussed in public.
Texts generated by language models are particularly laborious to review. They appear plausible at first glance: correct terminology, clean structure, specific line references. Whether there is a real defect behind them often only becomes clear after several minutes or hours of analysis. It is precisely this asymmetry – seconds to generate, hours to refute – that is the core of the problem.
Quality assurance is not purely a project matter
For companies, this initially looks like an internal affair of other people's projects. In fact, it affects their own supply chain. Components such as Curl are embedded in countless applications, containers, appliances and embedded devices. If a project's ability to process genuine security reports promptly suffers, the response time deteriorates for everyone who builds on it.
There is also a side effect that has so far received little discussion: if projects restrict reporting channels or introduce hurdles, the rules of the game also change for companies that want to submit findings themselves. Anyone reporting a vulnerability in future may have to provide more evidence, reproduction steps or proof of real-world impact.
What can be derived for one's own process
The lesson from the Curl case concerns less the question of whether AI is used in day-to-day development than how its output is handled. A few points can be transferred directly to internal workflows:
- Responsibility remains with humans. Anyone who passes on an AI-generated analysis – internally to the development team or externally to a project – should have verified it themselves beforehand. A report without reproduction is a supposition, not a finding.
- Define triage levels. It helps to distinguish between reproduced defects, plausible indications and unverified scanner or model output. This labelling should be attached to the report, not kept in individual people's heads.
- Evidence instead of description. A minimal reproduction case, a test case or a log excerpt is more valuable than a detailed prose description. With AI output, it is exactly the opposite: prose is available in abundance, evidence is missing.
- Put contribution rules in writing. Even for internal repositories, it is worth recording whether and how AI-assisted issues, pull requests or security reports may be submitted and which details are mandatory.
Review capacity as a planning factor
In many organisations, the introduction of AI tools was justified with a productivity promise: more code, more tests, more analyses in the same time. The Curl case shows the flip side. If the generation side is accelerated but the review side remains unchanged, a backlog builds up. This applies to public projects just as much as to internal teams in which pull requests are created faster than they can be read.
In practice, this means: review capacity belongs in project planning – as a line item of its own, not as a residual quantity. Anyone introducing additional code generation should at the same time define who reviews, according to which criteria and with what time budget. Automated pre-filtering, such as mandatory tests or static analysis before human review, can help but is no substitute for expert assessment.
Using open source responsibly
Another aspect is the relationship with the projects themselves. Anyone who benefits commercially from open-source components has a vested interest in their maintainers remaining able to act. That can mean financial support, but also simply discipline in handling reports: better one well-evidenced report than ten unverified ones.
For web and software projects, this leads to a clear consequence: AI may provide suggestions, but a human must still provide the proof – internally as well as towards third parties. Anyone who now defines review and triage processes cleanly and honestly plans for review capacity avoids the gained speed reappearing elsewhere as a backlog. And anyone who relies on open-source building blocks in their own supply chain should in future factor in their resilience as a risk, not just their range of functions.