Plan A · AI verification

AI 2040 Plan A — How You Can Help

Concrete next steps for supporting the R&D and deployment of AI verification technologies — whoever you are.

This page covers Plan A, the AI Futures Project’s proposal for a positive vision for AI development. In AI 2040 (Plan A), they propose a verified slowdown in AI development. This is one example of a use for AI verification, but it is far from the only approach the AI verification field is working on. Put simply, the AI verification field works on a toolbox of mechanisms which allow AI developers to make trustworthy claims about the systems they are developing and deploying. Sometimes the claims are useful for international deals; sometimes they support consumer or regulator trust.

Plan A is the AI Futures Project’s proposal for a positive vision for AI development. Under their proposal, in 2029, the US President announces that the US will pursue international cooperation to avoid an imminent intelligence explosion.

“The US and China want to not do an intelligence explosion in the near future. In order for an agreement along these lines to be trustless, they need verification infrastructure in place to make sure that AI development isn’t proceeding too quickly. They arrive at a simple solution for enforcing this: both sides agree to a temporary pause on new training runs.

A pause on training can be enforced by ensuring that compute is being spent on inference, which can be enforced by retrofitting existing data centres with inference-only devices. (Additional layers of enforcement include software-only privacy-preserving monitoring, and of course, old-fashioned human auditors.)”
— AI 2040

This description of 2029 requires AI verification technologies. Systems that provide sufficient assurance to both the US and China that defection by the other would either be caught or inconsequential.

On this page, we outline some of the things that you can do to support the R&D and deployment of these technologies.

If you’re interested in supporting AI verification work and would like further recommendations on how you can accelerate AI verification progress, then please fill in this expression of interest.

Choose your role

Which one are you?

AI Lab employee or leader

A lot of the key information about how frontier models are developed and deployed is housed within the AI labs; this makes them a natural place for AI verification work to take place.

Plan A Verification

Kick-start coordinated work on verification technologies and standards:

Initiate industry conversations on Plan A verification techniques. One option would be to use the Frontier Model Forum as a place to start inter-lab conversations.
Publicly announce the intention to work on AI Verification. Even simple public statements can encourage and accelerate other people’s AI verification work.
Run an external grant or venture funding programme to accelerate the AI verification field. Less than $15M has been spent on AI Verification technologies; even small increases in the available funding will accelerate the field. AI labs have a lot more talent reach than some other funders, and such a funding round would be a call to action for top talent.

AI labs can directly support technical AI verification R&D, both by running internal projects and by working with outside research groups.

Create an initial internal demo of inference verification (e.g. re-implement this demo on your stack with relevant frontier models)
Start an AI verification research track inside your lab. Research projects could be pulled from the list of technologies described in Plan A.
Give feedback on technical plans being developed by outside research groups.
Support joint verification experiments, including AI verification red-teaming exercises.

Ultimately, Plan A will require the cooperation of AI model developers, and the engineers and researchers at AI labs are likely to be involved in implementation. Gaining familiarity with and accelerating the tech readiness of AI verification technologies will help Plan A succeed.

Further Verification projects

Any verification project should start with the question: what claim are we trying to verify? An initial exercise for any AI model developer could be to answer these questions internally:

What existing commitments from AI developers (RSP, Frontier Safety Framework, Preparedness Framework) would be useful to verify?
What would reassure one AI developer that another developer is following best safety practices1?
What information can be unilaterally disclosed now that might support future high-stakes coordination?
What training and deployment practices should be considered “unsafe”?
What capability thresholds might be worth gating?

Working through these questions and, if appropriate, releasing the answers publicly would help start to give direction for technical AI verification projects and near-term verification goals.

If any of this is of interest, email us. We can help fill out these recommendations in more detail and make them more suitable to you.

“But to the extent that other relevant AI developers … prioritize safety and invest in legible demonstrations that they are doing so—as we intend to—commitments like this may help avoid an inadvertent “race to the bottom” on safety.” Anthropic RSP — Appendix A ↩

Government

In Plan A, the American President announces that the US will pursue international cooperation to avoid an imminent intelligence explosion. At this point, the government apparatus will mobilise to assess and then support the implementation of AI verification technologies. In the AI Futures Project’s scenario, governments start being interested in AI verification technology R&D in 2027.

The sooner work can start on AI verification, the better.

“Ultimately, governments will be involved in and have to trust and understand AI verification technologies. This means that either governments will need to become familiar with the technology very quickly at crunch time, or they should start preparing now.

On top of the necessity for governments to be involved in the implementation of AI verification, governments are also well placed to directly run AI verification R&D projects (due to both funding and security expertise).”
— AI 2040

R&D funding. Run an ambitious programme to develop open AI verification systems. Funding must be deployed very quickly, and the verification technologies must not end up tied up in the intelligence services. It is important that AI verification technologies are transparent and internationally trusted.
- Identify companies and research groups in your country that could be important contributors to AI verification technologies and convene them to understand what they need in order to start work on AI verification.
- Create verification demos deployed on national computing assets. Large data centres are not always available for test exercises, but national research computers can be suitable places for research deployments. Ideally create dedicated test ranges like the Side Channel Cloud or the UK’s Scaling Inference Lab.
- Experiment with leveraged funding models like Advanced Market Commitments for verification technologies.
International coordination. Where appropriate, raise AI verification during international dialogues.
- At the AI Summit in Switzerland in 2027, host initial dialogues on the role of the international community in developing and deploying AI verification technologies.
- Run a joint international AI verification experiment. Inspired by the Joint Verification Experiment, a US-Soviet collaboration to measure the explosive yields of nuclear tests, run an international exercise. This should test both the robustness of technologies plus encourage collaboration and standards building on how AI verification should be implemented.
Start a government AI verification programme. Some questions are best answered by governments. Plan A points out some of the roles that governments will have in tracking compute and purchasing network TAPs. Example areas of research:
- Task relevant government agencies to track compute locations (both inside and outside your country)
- Research on identifying data centres using National Technical Means. This might include detecting covert data centres from satellite images and heat signatures.
- Task relevant government agencies with working out how to secure (and verify the security of) recomputation servers. Of particular concern is the integrity of these servers.

These have been general recommendations for what governments might do. There are further detailed recommendations for what AI Security Institutes, National Security agencies, Ministries of Foreign Affairs, etc. can specifically do to support Plan A.

The role of each country and each government department is different. Reach out to us, and we can help you work out the work that you should be doing in your context.

Philanthropists

AI verification is speculative. There is no existing market and little R&D funding from governments, leaving a clear gap for philanthropy to fill. As the salience of AI verification increases, we expect the funding landscape to change, but for now, philanthropy has an important role in kick-starting the technical field. Because the field is fairly new a lot of the early grant making may need to be active, with philanthropists actively seeking out potential founders to start projects.

Rather than assuming that good grant opportunities will come through the door, funders should try running grant rounds and programmes aimed at creating new projects and organisations. Example programmes could include:
- Incubators/accelerators – For example, follow the model of the Big if True Science accelerator run by Renaissance Philanthropy. Running a BiTS (or similar) cohort dedicated to AI verification would be a way to turn early ideas on AI verification into concrete R&D plans that are capable of absorbing further funding.
- Requests for Proposals – Whilst not as active as other grant-making approaches, RFPs let you set out some specific goals to potential grantees. If distributed to the right groups (e.g. FRCs, academia, early startups), the RFP itself can really shape the way people are thinking about the technical problems they are solving. See Longview Philanthropy’s HEMs RFP from 2025.
- ARPA-style funding – In the 1950-60s the US DoD ran an ARPA programme to monitor compliance with nuclear test ban treaties. Likewise, focused ARPA programmes can focus on quickly developing and testing new verification technologies, especially high-ambition, high-trust verification projects.
In addition to kick-starting the field using active grant-making, the role of philanthropy could be to fund opportunities that would otherwise fall between the cracks. Ideas for projects that could be funded:
- Run a large AI verification red-teaming event (e.g. at DefCon)
- Fund a large central project to “own” Plan A verification. Their goal should be to ensure AI verification technology can be deployed in early 2029. The remit would cover not just the R&D but also the stockpiling and planning for scale-out. Before the end of 2026, they should be capitalised with at least $50-150M, and as the project succeeds, they should expect to scale to absorb more funding.
- Fund the creation of new research organisations dedicated to solving Plan A technology R&D challenges. This could include FROs (e.g. through Convergent Research), FRCs (e.g. organisations like Amodo).
- Define a set of R&D projects that can be handed off to “lower context” research organisations. It won’t be possible to immediately scale aligned R&D projects, so some projects should be handed over to academia and contract research organisations. This should be seen as a shot on goal rather than a sure-fire bet.
Fund speculative high-risk R&D beyond Plan A. Plan A is one possible solution to AI verification, but there are other approaches that might be suitable. Funding the creation and scaling of projects working on software-only solutions, cryptographic techniques, etc., could be really valuable. We cover all of the technical projects mentioned in Plan A in this piece (link).

These high-level ideas might not be a useful starting point. We’d be happy to speak to you about ideas for where to start on AI verification grant-making and connect you to existing philanthropists in the space: Coefficient Giving and Longview Philanthropy.

Venture Funders

A year ago, we were telling VCs that there would be a valuable opportunity for companies working to secure physical hardware (especially data centres) against AI-enabled cybersecurity threats and that soon governments would be interested in verifying the development and deployment of advanced AI systems.

A year on, Mythos has shown us that AI does really change the cyber landscape, and governments have shown interest in verifying the deployment of dangerous AI systems. On top of this, Anthropic has indicated that they are working on “provable inference, a technique for reliably, provably ‘signing’ AI model outputs” and OpenAI has called for international coordination.

Assuming that AIs will continue to improve means updating your thesis to include technology bets that were previously weird. A few opportunities that we see:

Data centre security, especially technologies that help data centres reach SL5.

Physical data centre and server security:
- Tamper-evident compute (e.g. server racks with advanced intrusion detection)
- Tamper-evident cameras (e.g. at-scale up-to-date equivalents of IAEA cameras)
- Personnel monitoring including
Hardware supply chain security
- Scanning devices for hardware Trojans/implants (link)
- Detecting covert backdoors in silicon design
- Detecting whether a fabricated chip matches the approved design
Side-channel mitigations – develop solutions that mitigate the information that can be stolen from side channels. Technologies should be deployed at frontier AI data centre scale:
- TEMPEST-proof server trays including shield
- Wardens for detecting covert side-channel communications

Plan A points to a world where the security-efficiency trade-offs have changed. Current enterprise applications do not look for national state-level security. As this changes, so do the types of technologies that need to be developed:

Reproducible machine learning infrastructure, including a reproducible network stack and software stack for training and inference.
Inference only chips.
Secure co-processors capable of performing verification tasks.
Treaty-grade verification hardware and software
Verification technologies that can be sold to labs immediately. For example, the verification systems served by Lucid Computing.

If you’re a VC who wants to add hardware security or AI verification to your thesis, we’re happy to hop on a call and give you a bit more context on the field, active startups, and open problems.

Engineers, researchers, and scientists

You’re needed!

The field is young and currently talent-constrained. The highest priority is to bring in people who could run new AI verification projects, but on top of that, we need individual contributors too.

Join an existing org. Orgs are hiring for engineers, TPMs, researchers, policy specialists, and more:
- Amodo Design
- Lucid Computing
- RAND Corporation
- AI-labs (OpenAI, Anthropic, Google DeepMind)
- FLI
- SASH
Go to a conference. The next relevant conference is the AI Security Forum in Las Vegas in August. There have previously been dedicated conferences on AI Verification.
Solve an open research problem. See our SitRep on Plan A verification or the AI Security Forum’s list of projects.
Speak to funders. The current active funders in the space are Coefficient Giving and Longview Philanthropy. They can be useful nodes for
Join an accelerator. Look out for new accelerators from ERA, Coefficient Giving, Constellation Institute, Pivotal, and MATS.

If you’re interested in working in or starting a new org in AI verification, please reach out, and we’ll be sure to connect you with the right people.