Deep Dive · Open-Source AI · Release Governance
Open-source software gains its strength from publication. Frontier AI does not behave the same way, because published model weights are an active capability: copyable, modifiable, redistributable, and impossible to withdraw. The release decision is therefore the last point at which safety can be designed in, and the first point at which it becomes irreversible.
or email Jansen@j-law.de
Originally published in the CR online blog on 27 October 2026. Read the full analysis: A Risk-Based Approach to Open Source AI.
Free and open-source software is defined by access and permission: source code published under an open licence so that anyone may inspect, modify, and redistribute it. That model built most of the modern internet, and its benefits are real: lower cost, no vendor lock-in, open standards, and public scrutiny that patches vulnerabilities faster than closed review.
A language model is not inert text. Its capability lives in its weights, and an open-weight release hands every recipient a system that runs locally, without telemetry, without an endpoint, and without a kill switch. The gap between the strongest open-weight and closed models narrowed from roughly a year to a few months through early 2026. The policy question is no longer whether open models are good, but which capabilities may responsibly be published.
The article's case rests on documented events, not projections. Three examples carry the weight:
The asymmetry that matters for legal and procurement work: a provider can detect, disrupt and report misuse only while it controls the endpoint. On downloaded weights, provider-side telemetry and central kill switches are simply absent. Detection and attribution are what make any enforcement regime possible, and open weights remove both.
Licences cannot fence capability. The Open Source Definition, criteria 5 and 6, forbids discrimination against persons, groups, or fields of endeavour. A licence that says "open, but not for military or malicious use" is therefore not open source, and a licence that is not open source buys little: it operates only after a violation is discovered, evidenced, and litigated. Even a clean preliminary injunction takes weeks or months, by which time the weights have circulated permanently.
Local deployment is not the same as public redistribution. Data residency, confidentiality, latency, and vendor independence are legitimate reasons to run a model on your own hardware. None of them requires publishing the weights. A model can be delivered locally under closed distribution, encrypted execution inside a hardware trusted execution environment, or a non-redistribution agreement. Conflating the two is a category error, and it is the most common one in this debate.
Release decisions should weigh three things:
| Tier | Profile | Release mode |
|---|---|---|
| Low risk | No relevant impact on critical infrastructure, military capability, or democratic processes; no meaningful autonomy or self-replication | Full open weights |
| Medium risk | Notable dual-use capability; safeguards integrated but strippable in days to weeks by specialists | Selective release: open weights for some components or versions, staged or gated for others |
| High risk | Autonomous agency, self-replication, or offensive cyber and bio capability; safeguards removable in minutes | Vetted access, staged distribution, monitored APIs; open release only once protection justifies it |
The framework runs in both directions. It is a whitelist, not a ban: where collective benefit clearly outweighs risk, publish openly.
The rule has two halves, and the first is routinely skipped.
Open the safeguards. Protective tooling, evaluation harnesses, and security classifiers get stronger under public scrutiny and shared maintenance. Opening safety research costs nothing in risk, because a safeguard built into release architecture and access protocol is not exposed to weight-level removal.
Make publication safe. Where capability is dangerous, the release process itself must carry the safety burden: staged distribution, vetted researcher access, monitored APIs. This is not secrecy for its own sake. It is the recognition that after publication, the only remaining mitigation is that a stronger model eventually supersedes the released one, which raises the risk again rather than resolving it.
If safety determines whether a frontier release is responsible, competing laboratories have to agree on baselines, and that raises an antitrust question the article takes seriously rather than dodging.
The objection is fair: competitors agreeing to pace output resembles a cartel, and pacing is futile if unguarded open models simply overtake the paced ones. The answer is not to abandon coordination, which leaves a race to the bottom in which every lab accelerates for fear that rivals will cut corners first. The answer is to make cooperation transparent, legally binding and publicly supervised: evaluators chosen by governments rather than by labs, published testing standards, and a narrow antitrust safe harbour confined strictly to safety protocols. Safety protocols should never be hoarded as proprietary advantage.
The framework is no longer only academic.
Open-source security already solved a version of this problem. Debian pairs a default commitment to public fixes with a time-limited embargo (capped at two weeks) and a shared public defect registry. Mapped onto frontier AI, that suggests an open, community-run evaluation harness and a public defect registry, combined with staged, embargoed distribution until known vulnerabilities are mitigated. Weights cannot yet be patched in place like a Linux package. The collaborative core of open-source security can still be preserved while governing capabilities that cannot be recalled.
For providers, the release decision becomes an accountable gate with recorded safety criteria and external review before high-risk weights become irreversible public goods. For buyers, it means asking a different question in procurement: not "is this model open?" but "which tier is it in, and what evidence supports that classification?" For counsel, it means the contract cannot fix what the architecture has already published.
Does this argue for closing AI? No. It argues for earned openness. Low-risk systems should be fully open, and safeguards should be public.
Can a licence stop military use of open weights? Not effectively. Field-of-use restrictions are incompatible with the Open Source Definition, and enforcement begins only after irreversible disclosure.
Is running a model locally the same as publishing it? No. Local deployment is compatible with closed distribution, confidential computing, and non-redistribution terms.
Which tier applies to a general-purpose frontier model? Rarely the low tier. Unless harmful capabilities can be prevented at the data or architecture level before publication, general-purpose frontier models belong behind gated access.