The most important rule on this agenda is not yet a public rule.
On August 4, 2026, Reuters and The Washington Post reported that United States officials had communicated to artificial intelligence companies that open-weight models would be left out of the government's new voluntary assessment process for advanced attack capabilities cyber. Both reports are based on people familiar with a private meeting, not a framework published by the White House.
The official document available is Executive Order 14409, signed on June 2, 2026. It mandates the creation of classified benchmarks to identify “covered frontier models” and a voluntary mechanism by which developers could provide early access to the government. The text does not say that open-weight models are exempt, does not publicly define the capacity limit and does not present the final tests.
So the publishable fact at this point is narrower: two newsrooms reported a privately discussed exception; The overall architecture of the program is official, but the full criteria and exclusion of open-weight models have not yet appeared in a public text.
What the reports said
According to Reuters, representatives from Meta, Anthropic, Google, Nvidia and OpenAI participated in the discussion. Two sources stated that the administration would not submit open-weight models to voluntary tests, while five sources confirmed that the meeting was taking place on rules that have not yet been published. The White House press office did not respond to the agency's request for comment.
The Washington Post published a separate confirmation, also based on two people who spoke on condition of anonymity. The newspaper reported that the framework would be completed, that systems with open weights would receive an exception and that the government did not intend to release its complete version. The Post itself declares to maintain a content partnership with OpenAI, one of the companies mentioned in the discussion.
None of the reports presented the document provided to the companies, a transcript of the meeting or a public statement from the White House confirming the exception.
Confirmed, reported and not yet published
| Point | Situation on August 4, 2026 | Base available | What cannot be concluded |
|---|---|---|---|
| The government has created a process to evaluate advanced AI cyber capabilities | Confirmed | Executive Order 14409 mandates ranked benchmarking and a voluntary framework for covered frontier models. | The text does not reveal which models meet the limit or how the tests will be scored. |
| Developers may offer early access to the government | Confirmed | The order provides access for up to 30 days before release to other partners considered trustworthy. | There is no general obligation for licensing or prior authorization in the public text. |
| Open-weight models will be left out of this process | Reported by Reuters and Washington Post | Both publications attribute the information to knowledgeable sources from a private meeting. | The exception is not yet confirmed by rule, official note or public framework. |
| Meta, Anthropic, Google, Nvidia and OpenAI participated in the discussion | Reported | Reuters cites five sources familiar with the meeting. | No minutes, individual positions or commitments made by each company were published. |
| The final framework is complete | Reported | The Washington Post attributes this information to an administration official. | Unable to examine the text, its versions, attachments or methodology. |
| Capacity limit, benchmarks, and rules for hybrid cases | Unpublished | The order says that benchmarking will be classified and that the limit will be determined by security authorities. | It is not known how open, closed, tweaked or models offered under different licenses will compare. |
Open-weight is not synonymous with open source
The distinction is decisive because “open”, “open-weight”, “open source” and “frontier model” describe different characteristics.
| Type or classification | What does it mean | What might be available | Relationship with security |
|---|---|---|---|
| Open-weight | Learned parameters, or weights, can be obtained and executed by third parties. | Weights and depending on the project, architecture, inference code or documentation. | Does not prove security or insecurity. Makes auditing and adaptation easier, but also reduces central control after distribution. |
| Open source | In addition to providing components, it grants freedom to use, study, modify and share the system. | As defined by the Open Source Initiative, must include adequate information about the data, training and execution code, and parameters. | Legal and technical disclosure is not a substitute for capacity testing, misuse assessment, or deployment security. |
| Closed or proprietary | The vendor keeps the weights in their control and typically offers the model per service, application, or API. | User interface, documentation and possibly evaluation reports. | The provider can monitor, change safeguards, or suspend access, but external researchers may have fewer elements to conduct independent audits. |
| Border Model | Model located at the front line of relevant capacity. It is a performance or risk rating, not a license rating. | Can be closed, open-weight or distributed via restricted access. | The key is what the model can do and under what conditions, not just how it is licensed. |
In other words, open-weight and closed are distribution and control categories; “frontier” is a capabilitycategory. An open-weight model can reach the boundary, while a closed model can remain below it.
The official terminology itself is not entirely uniform. The White House AI Action Plan, published in July 2025, grouped “open-source” and “open-weight” together in advocating an environment conducive to open models. The Open Source Initiative definition, however, requires more than simply publishing weights.
Why the possible exception matters
Testing may create different incentives for similar models
If two models have comparable cyber capabilities, but only the closed system is subject to the government process, the type of distribution will determine part of the regulatory treatment.
A possible interpretation is that the exception works as an incentive for North American companies to launch models with available weights. The White House Action Plan links these models to competition, academic research, local use of sensitive data and reduced dependence on closed suppliers. The Washington Post also pointed out that the exemption could favor United States investments in open alternatives to foreign models. This justification, however, was not officially presented as the reason for the exception.
Excluding from the framework does not mean excluding from all public evaluation
The North American government has already evaluated open-weight models by other programs. In April 2026, for example, the Center for AI Standards and Innovation, linked to NIST, tested DeepSeek V4 Pro in cybersecurity, software engineering, science, reasoning and mathematics tasks.
This indicates that the controversy is not about a blanket ban on testing open weights. The question is whether they will be left out of this specific pre-screening and government access mechanism created by Executive Order 14409.
Distribution changes the type of mitigation available
In a closed service, the company can modify filters, limit accounts, observe usage patterns, or deactivate a version. When weights are distributed, copies can be executed and modified outside of the original infrastructure, even without safeguards applied by the developer.
Affordable weights, on the other hand, allow universities, companies and researchers to perform independent assessments, make local adjustments and investigate behaviors without relying exclusively on the supplier.
Therefore, classifying all open-weight models as safe would be as inappropriate as treating them as inevitably dangerous. An NTIA report concluded that the evidence then available supported neither immediate categorical restrictions nor the promise that they would never be necessary. The recommendation was to monitor capabilities and benefits, assess thresholds of concern and preserve the possibility of proportionate actions.
What the government actually published
Executive Order 14409 establishes three main elements:
- Classified Benchmarking: Federal agencies must assess advanced cyber capabilities and set the threshold for a “covered frontier model.”
- Voluntary participation: Developers will be able to consult with the government on models under development and provide early access under confidentiality and security obligations.
- Absence of mandatory licensing: The order states that it does not create authorization for licensing, pre-approval, or mandatory government permission to launch models.
The public text does not say whether “voluntary” will only mean spontaneous adherence or whether participation could become relevant for public contracts, access to reliable partners, exports or other government decisions. It also does not explain the treatment of foreign models, versions adjusted by third parties or systems whose developer discloses the weights but restricts certain uses by license.
The questions the public framework would need to answer
A technically comparable policy would need to clarify, at a minimum:
- what cyber capability makes a model cross the coverage threshold;
- whether the criteria considers only the base model or also tools, agents, internet access and computational resources;
- why would the form of distribution justify inclusion or exclusion;
- how open-weight models that outperform closed systems in a relevant benchmark will be treated;
- what measures can correct a worrying result;
- who may contest, review or reproduce the evaluation;
- which results can be released without revealing sensitive benchmarks;
- whether foreign models available in the United States will receive equivalent treatment.
Five Democratic senators — Kirsten Gillibrand, Adam Schiff, Mark Warner, Christopher Coons, and Mark Kelly — sent a letter dated August 3, 2026 to administration officials calling for a public, durable, and technically sound framework. The letter also requests clarification on criteria, legal authorities, participation of independent experts, resources for affected companies, and consistent treatment among developers. Reports published the document on August 4th.
What to watch for in upcoming updates
The eventual publication of an official text could change important parts of this analysis. The most relevant signs will be:
- full or partial disclosure of the framework;
- public definition of “covered frontier model”;
- formal confirmation of exception for open-weight;
- demonstration from the White House or companies present;
- documentation on testing, fixes and early access;
- rules for foreign models and modified versions;
- Congressional proposals to transform the voluntary mechanism into law.
Until that happens, the most rigorous formulation is not “the United States has decided that open models are safe.” It is: anonymous sources told two publications that the government intends to leave open-weight models out of a voluntary process whose final design remains without public scrutiny.
Save the term table and evidence comparison. They should be updated when the framework, an official confirmation or new criteria are released.
