AI Slop and the CRA: What ISC and NLnet Labs Told the EU

A rumour is going around that the people behind BIND and Unbound lobbied Brussels to ban AI. They did not. What they did was put a sentence on the record that every team running an AI agent against open-source code should read. This post sorts out which document says what, quotes the part about LLMs, and turns it into a concrete bar for AI-generated issues, vulnerability reports and patches.
Two Documents, Often Mixed Up
The 2023 letter to Parliament. On 19 April 2023, ISC and NLnet Labs sent an open plea on the Cyber Resilience Act (CRA) and the Product Liability Directive to the ITRE committee's rapporteurs and shadow rapporteurs on the CRA. Their concern was liability and scope. Draft wording suggested that charging for support could make a non-profit open-source developer a commercial actor under the act, and the authors said their DNS and DHCP software appeared to fall into the most heavily regulated category. They asked Parliament not to jeopardise the non-profit funding model that pays for maintenance. The letter does not mention AI.
The 2026 joint response to the Commission. On 3 February 2026, CZ.NIC, ISC, NLnet Labs and NetDEF (the FRRouting maintainers) published a joint response to the European Open Digital Ecosystem Strategy call for evidence. It is about funding: multi-year grants for maintenance, downstream users paying for what they depend on, and a procurement norm with a budget line for upstream maintenance. It also asks the Commission to use hooks already in the CRA. These are the voluntary security attestation programme, the Article 13(5) due diligence duty on manufacturers that integrate third-party components, the Union-wide dependency assessment, and mandatory SBOMs with stable component identifiers.
The AI paragraph is in this second document, under the heading "AI technology will not replace human maintainers":
As of this writing, we have found LLM-based tooling to be more effective at wasting maintainer time (bogus vulnerability reports, nonsensical code contributions by well-intentioned individuals who do not understand what these algorithms produce), than at alleviating the bottleneck of core maintenance.
The paragraph goes on to say that even if LLM review tools become significantly useful, projects will still need humans to direct the tools, interact with the user community and take responsibility for the project. That is not a call for a ban. It is an argument that AI does not remove the need to fund human maintainers, and a warning about what low-effort AI output costs them.
The Cost Is Measurable
The DNS maintainers are not alone. The curl project ended its bug bounty on 31 January 2026. Over the program's life it confirmed 87 vulnerabilities and paid more than 100,000 USD. Daniel Stenberg wrote that the confirmed rate was above 15% in earlier years and fell below 5% from 2025. He blamed an explosion in AI slop reports, combined with lower quality in other reports and more people poking holes rather than helping. He said the submissions "take a serious mental toll to manage and sometimes also a long time to debunk."
A bogus report can cost the reporter seconds and the maintainer hours, because a security report cannot be ignored until someone proves it wrong. Volume is not the main problem. The cost lands on the side that never asked for the work.
AI-found bugs can be real and valuable. We covered Anthropic's Firefox engagement, where Mozilla validated the findings and shipped fixes. The difference is not the model. It is whether a human did the verification before the maintainer had to.
Where the CRA Comes In
The Cyber Resilience Act is Regulation (EU) 2024/2847. It entered into force on 10 December 2024. Its reporting obligations have applied since 11 September 2026, and the main obligations apply from 11 December 2027. Three parts of the regulation text matter for anyone pointing AI tools at code they ship:
- Article 13(6). A manufacturer that finds a vulnerability in an integrated component, including an open-source one, must report it to whoever maintains that component. If the manufacturer develops a fix, it must share the code or documentation with the maintainer. So when your agent flags a bug in a library you ship, an upstream report is not optional goodwill. It is part of your obligations, and its quality is your responsibility.
- Article 14. Actively exploited vulnerabilities go to the coordinating CSIRT and ENISA through the single reporting platform. The deadlines are an early warning within 24 hours, a notification within 72 hours, and a final report within 14 days after a fix is available. A tool that labels a finding "actively exploited" without evidence starts a clock you then have to justify or withdraw.
- Stewards and volunteers. The Commission's CRA open-source page says the act does not apply to developers who contribute code to projects that are not under their responsibility, and that non-monetised free software is not a commercial activity. Open-source software stewards, such as foundations that support a project on a sustained basis, have lighter duties and are not subject to administrative fines under Article 64(10).
Taken together, the CRA puts responsibility on the manufacturer that ships the product. The maintainer gets the reports. If an AI tool in your pipeline generates those reports, your organisation owns their accuracy. This post is not legal advice, but the direction is clear enough to design workflows around.
The Bar: What an AI-Generated Artifact Must Contain
Use this as a gate. If an artifact is missing a required field, it stays inside your organisation.
| Artifact | Must contain before it leaves your team | Channel | If a field is missing |
|---|---|---|---|
| Bug issue | Exact version or commit, minimal reproduction, observed vs expected output, search of existing issues | Public tracker, per CONTRIBUTING | Keep it internal |
| Vulnerability report | Reproducer against unmodified upstream, sanitizer or crash trace, attacker preconditions, reachable code path, named human contact | Private channel from SECURITY.md | Do not send; file an internal ticket |
| CVE request | Maintainer agreement that it is a vulnerability, affected version range, fixed version or mitigation | Project's CNA or its process | Wait for the maintainer |
| Patch | Minimal diff, a test that fails before and passes after, passing project CI, a sender who can explain every line | Project's review flow | Send an issue instead, or nothing |
| Internal advisory | Evidence links, confidence grade, which claims a human checked | Your incident or security channel | Mark it unverified |
Four rules sit on top of the table:
- A named human signs it. Someone who read the output, ran the reproducer and can answer follow-up questions without pasting them back into the model.
- Disclose the tooling. Say which parts the AI produced. Maintainers triage faster when they know what they are reading.
- Severity comes from preconditions, not adjectives. "Remote code execution" without a reachable path from untrusted input is severity inflation that the maintainer has to disprove.
- Check the docs first. Rule out documented behaviour, debug-only code and configuration the project already warns against.
Failure Modes to Catch Before Sending
Each of these patterns is cheap to test for on your side and expensive to debunk on theirs:
- Unreachable sink. The model finds a risky function call but no input path reaches it in a supported build.
- Invented API. The report names functions, flags or files that do not exist in the stated version.
- Warning-silencing patch. The diff makes a static analyser quiet and changes behaviour nobody asked to change.
- Reply relay. The maintainer asks a question and the answer is more model output, not a test run.
- Wrong channel. A plausible security issue lands in a public tracker, which hurts the project even if the bug is real.
What This Means for SRE Agents
An AI SRE agent produces the same artifact types: findings, root-cause claims, proposed fixes and tickets. The maintainer of your production systems is your on-call engineer, and they pay the same debunking cost when an agent states a guess as a fact.
Akmatori is built around that constraint. In its architecture, memories carry evidence grades: an agent's causal claim is always unverified, and only an operator can mark a memory confirmed. The run log of each incident shows the agent's tool calls with their arguments and output, so a reviewer can check a claim against the evidence. Per the integrations docs, risky actions become proposals that wait for approval. The Jira connector cannot comment on or create issues until you turn on Allow write operations in its settings. Applied to upstream work, the same pattern holds: let the agent find and draft, and let a human verify, sign and send.
Conclusion
The DNS maintainers did not ask Europe to ban AI. They asked the EU to fund maintenance, use the CRA to make manufacturers care where their components come from, and remember that maintainer time is the scarce resource. If your tools generate reports for other people's code, meet the bar above before you hit send.
If your team wants AI-assisted incident workflows on your own infrastructure, Akmatori helps SRE teams investigate alerts, coordinate response, and automate safe infrastructure actions with evidence a human can check. Powered by Gcore for global infrastructure reliability.
