AI Forge: DARPA and NSF Just Built a Direct Pipeline From University Labs to National Security AI Research. The June 22 RFI Decides Who Gets In.

For about a decade, AI safety and security research at U.S. universities has lived in a strange middle space. The frontier labs — OpenAI, Anthropic, Google DeepMind, Meta — control the models, the compute, and the empirical questions that matter most. Defense agencies fund a great deal of AI, but very little of it flows to the parts of academia working on interpretability, control, and adversarial robustness in a way that mirrors how industry actually deploys these systems. NSF runs a respectable program for trustworthy AI through Secure and Trustworthy Cyberspace and its smaller alignment efforts. DARPA runs its own threads through the Information Innovation Office. The two have rarely synchronized, and neither has moved at the pace of the technology.

That changed on June 1, 2026, when NSF and DARPA jointly announced AI Forge — a co-governed program designed to fund university-led research on the AI problems the national security community considers most pressing. The structure is unusual: a forum administered by an as-yet-unnamed nonprofit, governed jointly by DARPA, NSF, and the Center for AI Standards and Innovation at NIST, designed to issue multiple awards annually in the $750,000 to $3 million range for projects up to one year long. The forum launches in summer 2026. Before it does, the agencies want to know which university groups can actually do this work — and they are asking through a Request for Information posted at sam.gov as DARPA-SN-26-80, with responses due June 22, 2026.

For grant-funded AI researchers, the RFI is not a funding opportunity. It is the gate that decides which groups will be invited to compete when the real solicitations open. If your lab does not respond, the program leads will not know you exist when they design the first round of calls. That is a significant — and surprisingly literal — consequence for a single sam.gov form.

This piece works through what AI Forge actually is, what the three thrust areas mean in practice, what the RFI is asking universities to do, and how to position a response that puts your group on the inside of the program rather than on the outside reading about it on Twitter.

What AI Forge Actually Is

The announcement frames AI Forge as a "jointly governed forum that will fund, guide, and manage university-led research on AI interpretability, AI control, and adversarial robustness." The administrative shell is a nonprofit — the agencies have not yet named the operator — that will sit between the federal sponsors and the universities. The forum will be co-governed by DARPA, NSF, and CAISI, the NIST center created in 2025 to handle AI evaluation and standards.

The choice of a nonprofit forum, rather than a normal BAA or NSF program, is deliberate. Three reasons stand out.

First, speed. DARPA's standard BAA cycle takes months from solicitation to award. NSF's review cycles can take longer. The frontier AI field moves fast enough that a 2024 alignment research question can be obsolete by 2025. The forum model is meant to compress decision cycles to weeks, not quarters.

Second, portfolio coherence. AI safety research has historically been fragmented across small grants from different agencies, none of which were coordinating on what the field actually needs. By concentrating funding in a single forum with shared governance, the agencies hope to commission coherent portfolios — say, twenty interpretability projects designed to compose rather than overlap.

Third, industry-academia coupling. The forum is designed to "enable a more robust exchange of talent and ideas across universities, frontier AI companies, and government than is possible today." Translation: the forum expects industry researchers to flow in and out of academic projects, share models and benchmarks, and effectively treat university PIs as part of the same technical workforce as Anthropic or DeepMind employees, at least while the projects run.

Award sizes are calibrated to that goal. A $750K-to-$3M, one-year project is large enough to fund a postdoc, several graduate students, compute, and travel — but small enough to be one of many awards per year, not a flagship program that captures all the air in the room. Multiple awards annually means the forum is functionally a rolling solicitation engine, not a one-shot competition.

The Three Thrust Areas, Translated

The accompanying NSF-DARPA report identifies fifteen research challenges spread across three thrust areas. The thrusts are stated abstractly in the public materials. Here is what they mean if you are writing an RFI response.

Interpretability

"Interpretability" in the AI Forge framing is not academic visualization work on small CNNs. It is the family of techniques — mechanistic interpretability, sparse autoencoders, circuit-level analysis, concept probing, activation steering — that lets a researcher inspect why a frontier-scale model produced a specific output. The national security interest is straightforward: a model used in a high-stakes decision needs to be inspectable in a way that goes beyond "it scored well on a benchmark."

If your group works on interpretability, the RFI is asking you to be specific about which scale of model you have analyzed, which techniques you have applied, what infrastructure you bring (e.g., model weights access agreements with frontier labs, your own training runs at meaningful scale, novel tooling), and what questions you can answer that the field cannot. Generic "we work on explainability" responses will not differentiate.

Control

"Control" is the harder bucket to translate because the term has shifted meaning over the last 18 months. In the current AI safety vocabulary, control refers to ensuring that AI systems do what their operators intend even when the systems are highly capable and operate over long horizons — the territory that overlaps with alignment, oversight, scalable supervision, and red-teaming, but with an engineering emphasis on demonstrable assurance rather than philosophical guarantees.

For RFI purposes, control work that maps onto national security includes: techniques for verifying that an AI agent's actions match a specification, methods for catching deception or sandbagging in evaluations, scalable oversight protocols for tasks human reviewers cannot grade directly, and runtime monitoring of multi-agent systems. If your group has a paper trail in any of these areas, this is your thrust.

Adversarial Robustness

This is the most familiar bucket, and the one with the deepest academic literature. Adversarial robustness covers everything from input perturbation attacks on vision and language models, to jailbreak resistance, to defenses against poisoning, to robustness under distribution shift in deployment environments. The national security framing pushes the field toward contested deployment: assume your model is being attacked by sophisticated adversaries with access to gradients, weights, or rapid query budgets, and design defenses that hold.

The RFI wants to know which adversary models you can defeat, which you cannot, and what testbeds you have built. "We've published on adversarial examples" is too generic. "We maintain a benchmark for X attack class against Y model family with Z baseline defenses" is the level of specificity that moves a group up the list.

What the RFI Is Actually Asking

The RFI itself is light on prescriptive structure — deliberately so, because the program leads want to see what universities will volunteer. Based on the public language, responses should cover three things.

Capabilities. What can your group do that the field cannot? Tooling, infrastructure, model access, datasets, prior results. The honest version of this is the one that gets read.

People. Who specifically is doing the work, and what is their availability over the next 12 to 24 months? AI Forge expects to fund projects that begin shortly after the forum launches in summer 2026, which means the program needs PIs who can start in fall 2026, not whenever a normal grant cycle would permit.

Couplings. Who do you already work with in industry and government, and can you bring those relationships to a forum project? The forum is built around exchange of talent and ideas; groups that already have those couplings will be lower-friction picks than groups that would need them constructed from scratch.

What the RFI does not promise: a slot in the funded portfolio. The repository the agencies are building will inform the design of solicitations, but the eventual funding decisions will run through whatever competitive mechanism the forum nonprofit ends up using. Still, being in the repository is functionally a prerequisite. Groups that skip the RFI risk discovering, six months from now, that the first round of calls was written around questions their lab could have shaped.

How This Reshapes the AI Safety Funding Map

Three structural shifts deserve attention.

First, AI Forge consolidates demand at the small end of the academic AI safety market. NSF's SaTC and assorted small-dollar programs have historically been the main entry point for university interpretability and robustness work. With AI Forge offering $750K-to-$3M awards in fast cycles, the small-dollar programs become farm leagues — places to build a record before competing for forum funding. That has implications for early-career PIs deciding which solicitations to write to in 2026 and 2027.

Second, DARPA and NSF are publicly admitting they have not been moving fast enough on these problems. The forum structure exists because the standard BAA-and-program-officer model is not delivering, and the agencies needed a vehicle that could keep up. That is a candid institutional admission. It also signals that future agency programs in AI safety are likely to adopt similar forum or consortium structures rather than the standard format.

Third, CAISI's involvement is a tell. The NIST AI evaluation center being a co-governor of the forum suggests that AI Forge-funded research will feed directly into national AI evaluation standards. Groups whose interpretability or robustness work informs CAISI benchmarks will have outsized policy influence in addition to the grant funding — a double benefit that does not exist with standalone NSF or DARPA programs today.

The Eleven Days Before June 22

If your group could plausibly compete for AI Forge funding, the next eleven days matter. The RFI is short and forgiving on format, but the substantive bar is real. A weak response — one that reads like a recycled NSF biosketch with vague claims about working on "trustworthy AI" — will land your lab on the repository but at a position that does not influence solicitation design. A strong response that names specific capabilities, specific people, and specific industry-government couplings will put you in the conversation when the first round of forum funding is scoped.

The mechanical bar is low: read DARPA-SN-26-80 at sam.gov, follow the submission instructions, and respond before June 22, 2026. The strategic bar is the one that matters. Treat this as the moment when the agencies decide which universities are actually in the network they are building — because that is what it is.

For Granted readers tracking AI safety funding more broadly, this is the most consequential agency move of Q2 2026. The interpretability and control fields have been waiting for federal funding that matches the scale and pace of the problem. AI Forge is the first serious attempt. Whether it succeeds will depend on which universities show up in the next two weeks.