DEV Community

Cover image for Why Open Source Still Wins Trust in the AI Era, Even as Attacks Go Automated
Mir Arshad Ali Talpur
Mir Arshad Ali Talpur

Posted on

Why Open Source Still Wins Trust in the AI Era, Even as Attacks Go Automated

There is a tempting argument going around. Attackers now use AI to find and exploit weaknesses at machine speed, public code is easy to study, so the safe move is to close the code and hide behind the door.

We think that argument is wrong, and this article explains why. Our view is simple: in the AI era, the vendors who keep their code open and back it with verifiable evidence will win more trust than the vendors who ask buyers to take their word. Automated attacks are real, and the recent evidence is sobering. But the evidence points at weak pipelines, overloaded maintainers and unverified dependencies, not at openness itself. The answer to those problems is more verification, not less visibility.

This is an advocacy piece, and we say so. We build ZizkaDB, an open source causal audit database for AI agents, and we chose open source deliberately. We will make the case for it with recent facts, give the strongest counterpoint fairly, and finish with what it means for ZizkaDB.

A note on sources. The facts below come from public reports, security research notes, legal commentary and surveys published in 2026. Some are from vendors or blogs with commercial interests, and we flag that where it matters. Figures differ between sources, so we attribute them or give approximate values. Nothing here is legal advice.

The threat is real, and we should look at it squarely

Supply chain attacks now propagate by themselves. A Cloud Security Alliance research note describes a self-propagating npm worm, dubbed Mini Shai-Hulud, attributed to a group called TeamPCP. According to the note, a wave on 11 May 2026 published hundreds of malicious package versions across well over a hundred npm packages, plus two on PyPI, in under six hours, and reached AI developer ecosystems including TanStack, Mistral AI, UiPath, OpenSearch and Guardrails AI. Two versions of the note give slightly different counts, such as 404 malicious versions across 172 npm packages in one and 373 versions across 169 packages in another, so treat the exact numbers as approximate.

The note reports that the worm produced packages carrying valid SLSA Build Level 3 provenance attestations, by hijacking the legitimate release pipeline instead of stealing credentials and publishing from outside. A control many teams rely on said the package was fine. A separate vendor analysis reports that the worm's source was published on 12 May and that derivative campaigns followed within weeks.

Discovery is being industrialised, and maintainers feel it first. Help Net Security reported in May 2026 that AI-powered industrialisation of vulnerability discovery is a much bigger problem for open source projects than for big organisations, because open source relies on volunteer maintainers with limited time. Secondary accounts report that curl ended its paid bug bounty on 31 January 2026, that its confirmed-vulnerability rate fell from roughly 15 percent to under 5 percent as low-quality AI-generated reports flooded in, and that it paused vulnerability reports from 1 July to 3 August. Other reports say the Internet Bug Bounty paused new submissions in March, and that Linus Torvalds said in May that the kernel's security list had become almost entirely unmanageable because of duplicate AI-generated reports. Several of these come from blogs summarising primary statements, so check the projects' own announcements before quoting them.

These are not small problems. They are the honest cost side of open source in 2026, and the rest of this article does not wave them away.

Why the threat does not undercut the case for open source

First, closing the code does not remove the exposure. Black Duck's 2026 Open Source Security and Risk Analysis, covering 947 audited codebases, found that 97 percent of commercial codebases contain open source components, according to a commentary on the report, and that open source vulnerabilities per codebase doubled to 581. A closed product built on open components inherits every one of those risks. The worm above hit packages that closed and open products alike depend on. Keeping your own code private does not protect you from a poisoned dependency. It only prevents your customers from seeing what you depend on.

Second, the attacks target pipelines and people, not openness. The worm's novelty was hijacking a legitimate release pipeline. The curl story is about maintainer time. Neither is a flaw that disappears when the repository is made private. A closed vendor with a weak build pipeline and a small security team has the same problem, plus no outside reviewers to notice. The principle that security should not rest on secrecy is old and widely held. We offer it as an argument, not as data, but the recent evidence is consistent with it.

Third, open code lets buyers verify things that closed vendors can only assert. With open source, a buyer can read the code, build it, check what is running against what was published, self host it, and keep the data inside their own environment. When an attack like the worm hits, an open project can be inspected, its dependencies listed and its fix reviewed in public. With a closed product, the buyer waits for a statement.

Fourth, buyers are already telling us what they value. A Linux Foundation survey of 316 AI professionals found that 94 percent of organisations use generative AI, that about 41 percent of their AI code infrastructure is open source on average, and that 52 percent prioritise transparency and source code access when choosing AI solutions. A Red Hat survey of UK organisations on AI sovereignty found that four in five said open source offers greater control over how AI is built and where it runs, and 87 percent named transparency and easier auditability as the most valuable benefit for building trust in AI. These are survey results with their own sponsors and sample limits, but they point the same way.

Fifth, trust is the commercial advantage. The Linux Foundation's State of Commercial Open Source 2026 report, based on a survey of 225 founders and enterprise leaders, identifies community trust as the top competitive advantage for commercial open source companies, and enterprise sales capability as the main scaling bottleneck. The model works: win trust in the open, then serve enterprises that need it.

Sixth, regulation rewards traceable components. The EU Cyber Resilience Act's reporting obligations for manufacturers have applied since 11 September 2026, covering actively exploited vulnerabilities and severe incidents, including for products already on the market. Non-commercial open source is outside the Act, but commercial use of open source is not, so vendors must know what is inside their products. Open source stewards, who systematically support open source intended for commercial use, have lighter obligations from 11 December 2027. Buyers will ask vendors how components are tracked, and open projects with public dependency lists are better placed to answer.

The fair counterpoint

A fair reading also has to include the arguments on the other side. Critics say that open code is only as safe as the attention it gets, and in 2026 that attention is scarce. Most projects have very few maintainers, and the curl and kernel stories show them under strain. Visibility is not verification: anyone can read the code, but few buyers do, and the worm showed that even strong provenance controls can be subverted. Some security teams will also argue that for the highest assurance needs, a closed system under a vendor's contractual liability is simpler to govern.

These points deserve a serious response, and ours is this. If visibility alone is not enough, the answer is to add verification and to fund the work, not to close the door. Which leads to what open source vendors owe their buyers.

What open source vendors must do to earn the trust

Being open is the start, not the finish. Five commitments turn openness into trust.

Maintain a visible security process. A published policy, a way to report vulnerabilities, and a record of how fast fixes ship.
Harden the release pipeline. Treat the pipeline as part of the attack surface, since provenance alone can be subverted.
Track dependencies and publish what is inside. Keep an inventory, update promptly and be ready to say which versions are affected when a component is compromised.
Support self hosting and verification. Let customers build and check what runs against what was published, and keep their data in their own environment.
Give customers evidence they control. When something goes wrong, they should be able to reconstruct what happened from records they own, without waiting for the vendor.

Commercial open source companies also have a duty that individual hobby projects cannot meet: fund the maintenance. If a company profits from open code, it should pay for the security work that sustains it.

Where ZizkaDB fits

We chose open source for ZizkaDB for the reasons above, and we think a database whose job is to provide audit evidence should itself be open to audit.

Open code, inspectable by buyers. The ZizkaDB core is licensed AGPL-3.0 and the MCP server is MIT licensed. Teams can read the code, self host it in a single Docker container, and keep the audit record inside their own environment, which matters for sovereignty and data residency.

Verification, not just visibility. ZizkaDB records each agent step as an event, links it to the event that caused it, and protects the record with SHA-256 checksums and a per-agent hash chain, so edits to history are detectable. If you store the chain head outside the database, you can check on your own schedule that nobody rewrote the record. That is the fifth commitment above: evidence you control.

Evidence for when prevention fails. The threat evidence says compromises will happen. When one does, the useful questions are what ran, what it read and what it affected. If you log the tools, packages and MCP servers your agents call, with versions, and link decisions to the inputs behind them, the lineage and the impact route help scope which agent actions depended on a component later found to be bad. An illustrative record, with parameter names to be checked against the SDK reference:

python
call = await db.log(agent="ops-agent", event="tool_call",
data={"tool": "mcp-server-x", "version": "1.4.2", "target": "ticket-881"},
parent_id=decision.id)

A model that funds the work. Open source projects need sustainable funding. ZizkaDB offers a managed Cloud edition at db.zizka.ai for teams that want fleets, multi-agent and agent to agent recording, and that is how we aim to keep the open core maintained.

Our own posture, stated carefully. The ZizkaDB MCP server is listed on AddonAudit, which reports passing automated checks and a metadata score of 100 for public metadata. That page describes itself as an automated smoke test and not a security guarantee, and we say the same. We encourage buyers to run their own review of the code. That is the point of keeping it open.

Limits. ZizkaDB records, links and shows that the record is unchanged. It does not prevent supply chain attacks, detect malicious packages, or judge whether an output is correct, and it does not replace dependency scanning, signing, sandboxing or least privilege. The open source edition records one agent per install, and fleets and agent to agent recording use the Cloud edition. ZizkaDB supports record keeping and investigation and does not by itself make any system compliant with the EU AI Act, the Cyber Resilience Act or any other law. Confirm the specific features you rely on against the current documentation.

The takeaway

Automation has made attacks on shared code cheaper, and it has made buyers more demanding about what they run. Closing the code answers the first by weakening the second. We think the better answer is to keep the code open, harden the process, fund the maintainers and give buyers evidence they can check for themselves. Openness earns the first meeting. Verifiable evidence closes the deal.

If you run AI agents, start small. Pick one agent, log its tool calls with versions and the events that caused them, and verify the chain on a schedule. ZizkaDB runs self hosted in a single Docker container, with a managed Cloud edition at db.zizka.ai for multi-agent setups.

Corrections: if you represent a party named here and believe a statement is inaccurate, contact us and we will review it promptly.

Sources: Cloud Security Alliance research notes on Mini Shai-Hulud (May 2026); TuxCare analysis of 2026 supply chain attacks; Help Net Security, AI is drowning software maintainers in junk security reports (May 2026); Pinggy, curl's Summer of Bliss (July 2026); Falcao, AI bug reports and open source (May 2026); Resilient Cyber, Vulnpocalypse (2026 Black Duck OSSRA commentary); Linux Foundation survey of AI professionals, via CO/AI; Linux Foundation and Serena, The State of Commercial Open Source 2026; Red Hat UK AI sovereignty survey, via IT Brief UK; legal commentary on the EU Cyber Resilience Act reporting obligations from Goodwin, Travers Smith, Matheson and HLC (September 2026); ZizkaDB documentation.

Top comments (0)